Add Studio 2 caption jobs via Qwen2.5-VL GGUF on the host agent.

Exclusive llama-server load/unload on the 5080, queued Describe UI with resultText, Copy, and Use as prompt.

Co-authored-by: Cursor <cursoragent@cursor.com>

assets/css/display.css

@@ -0,0 +1,118 @@
+/* Studio 2 display skins — Classic | Studio. Layout tokens only; palette matches main.css. */
+
+:root,
+[data-display='classic'],
+[data-display='studio'] {
+  --display-bg: #09090b;
+  --display-fg: #e4e4e7;
+  --display-muted: #a1a1aa;
+  --display-border: rgba(255, 255, 255, 0.1);
+  --display-amber: #fbbf24;
+  --display-amber-fg: #09090b;
+  --display-radius: 8px;
+  --display-focus: rgba(252, 211, 77, 0.5);
+}
+
+[data-display='classic'] {
+  --chrome-gap: 0.5rem;
+  --rail-h: 2.25rem;
+  --card-pad: 0.75rem;
+  --display-type-title: 1.75rem;
+  --display-type-body: 0.875rem;
+  --display-viewer-radius: 0.75rem;
+}
+
+[data-display='studio'] {
+  --chrome-gap: 1rem;
+  --rail-h: 3rem;
+  --card-pad: 1rem;
+  --display-type-title: 1.875rem;
+  --display-type-body: 0.9375rem;
+  --display-viewer-radius: 8px;
+}
+
+[data-display] .display-title {
+  font-size: var(--display-type-title);
+}
+
+[data-display] .display-body {
+  font-size: var(--display-type-body);
+}
+
+[data-display] .display-stack {
+  display: flex;
+  flex-direction: column;
+  gap: var(--chrome-gap);
+}
+
+[data-display] .display-card {
+  border-radius: var(--display-radius);
+  border: 1px solid var(--display-border);
+  padding: var(--card-pad);
+}
+
+[data-display] .display-viewer {
+  border-radius: var(--display-viewer-radius);
+}
+
+[data-display] .display-rail {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 0.5rem;
+  min-height: var(--rail-h);
+}
+
+[data-display] .display-rail__seg {
+  min-height: var(--rail-h);
+  border-radius: var(--display-radius);
+  border: 1px solid var(--display-border);
+  padding: 0.5rem 1rem;
+  font-size: var(--display-type-body);
+  color: var(--display-fg);
+  background: transparent;
+  outline: none;
+  transition: border-color 0.15s ease, background-color 0.15s ease, color 0.15s ease;
+}
+
+[data-display] .display-rail__seg:focus-visible {
+  box-shadow: 0 0 0 2px var(--display-focus);
+}
+
+[data-display] .display-rail__seg[aria-pressed='true'],
+[data-display] .display-rail__seg[data-active='true'] {
+  border-color: var(--display-amber);
+  background: var(--display-amber);
+  color: var(--display-amber-fg);
+}
+
+[data-display] .display-rail__seg:disabled,
+[data-display] .display-rail__seg[aria-disabled='true'] {
+  opacity: 0.55;
+  cursor: not-allowed;
+}
+
+[data-display='studio'] .display-rail__seg {
+  font-weight: 600;
+  padding: 0.625rem 1.25rem;
+}
+
+[data-display='classic'] .display-rail__seg {
+  min-height: 2.25rem;
+  padding: 0.375rem 0.75rem;
+  font-weight: 500;
+}
+
+[data-display] .display-cta {
+  border-radius: var(--display-radius);
+  background: var(--display-amber);
+  color: var(--display-amber-fg);
+  font-weight: 600;
+}
+
+[data-display] .display-cta:focus-visible {
+  box-shadow: 0 0 0 2px var(--display-focus);
+}
+
+[data-display] .display-control {
+  border-radius: var(--display-radius);
+}

components/LoraSelect.vue

@@ -6,22 +6,21 @@
       <li
         v-for="(item, index) in model"
         :key="item.name"
-        class="rounded-xl border bg-zinc-950/80"
+        class="rounded-lg border border-[#2A2A2E] bg-[#1A1A1D] p-2.5"
         :class="[
-          compact ? 'px-2 py-1.5' : 'px-3 py-2',
-          dragOverIndex === index ? 'border-amber-300/50' : 'border-white/10',
+          dragOverIndex === index ? 'border-amber-300/50' : '',
           draggingIndex === index ? 'opacity-60' : ''
         ]"
         @dragover.prevent="onDragOver(index)"
         @drop.prevent="onDrop(index)"
         @dragleave="onDragLeave(index)"
       >
-        <div class="flex items-center justify-between gap-2">
-          <div class="flex min-w-0 items-center gap-2">
+        <div class="flex items-start justify-between gap-2">
+          <div class="flex min-w-0 flex-1 items-start gap-2">
             <button
               v-if="model.length > 1"
               type="button"
-              class="shrink-0 cursor-grab touch-none text-zinc-500 hover:text-zinc-300 active:cursor-grabbing disabled:cursor-not-allowed disabled:opacity-40"
+              class="mt-0.5 shrink-0 cursor-grab touch-none text-zinc-500 hover:text-zinc-300 active:cursor-grabbing disabled:cursor-not-allowed disabled:opacity-40"
               :disabled="disabled"
               draggable="true"
               title="Drag to reorder"
@@ -31,61 +30,91 @@
             >
               <span class="block leading-none" aria-hidden="true">⋮⋮</span>
             </button>
-            <span class="min-w-0"><span class="block truncate text-zinc-200">{{ labelFor(item.name) }}</span><span v-if="filenameSubtitle" class="block truncate text-[11px] text-zinc-500" :title="item.name">{{ item.name }}</span></span>
+            <span class="min-w-0 flex-1">
+              <span class="flex min-w-0 items-center gap-1.5">
+                <span class="block truncate font-medium text-[#F3F1EA]" :title="labelFor(item.name)">{{ labelFor(item.name) }}</span>
                 <LoraTriggerTip
                   :text="triggerFor(item.name)"
                   :aria-label="'Activation for ' + labelFor(item.name)"
                   @insert="emit('insert-trigger', $event)"
                   @save="saveTrigger(item.name, $event)"
                 />
+              </span>
+              <span v-if="filenameSubtitle" class="mt-0.5 block truncate text-[11px] text-[#9A9588]" :title="item.name">{{ stemFor(item.name) }}</span>
+            </span>
           </div>
           <button
             type="button"
-            class="shrink-0 text-[11px] text-zinc-500 hover:text-red-300 disabled:opacity-50"
+            class="shrink-0 text-[11px] text-[#9A9588] hover:text-red-300 disabled:opacity-50"
             :disabled="disabled"
             @click="removeAt(index)"
           >
             Remove
           </button>
         </div>
-        <div class="mt-1.5 flex flex-wrap gap-3">
-          <Thumbwheel
-            :model-value="item.strengthModel"
-            label="Model"
-            :lock-key="`lora-model:${item.name}`"
-            :min="LORA_STRENGTH_MIN"
-            :max="LORA_STRENGTH_MAX"
+
+        <div class="mt-2 space-y-1.5">
+          <label class="grid grid-cols-[3.25rem_minmax(0,1fr)_2.75rem] items-center gap-2 text-[11px] text-[#9A9588]">
+            <span>{{ clipStrength ? 'Model' : 'Strength' }}</span>
+            <input
+              type="range"
+              class="lora-strength-range min-w-0"
+              :min="0"
+              :max="2"
               :step="LORA_STRENGTH_STEP"
-            :fine-step="LORA_STRENGTH_STEP"
-            :coarse-step="0.25"
-            :digits="2"
-            compact
               :disabled="disabled"
-            @update:model-value="setStrength(index, 'strengthModel', $event)"
-          />
-          <Thumbwheel
-            :model-value="item.strengthClip"
-            label="CLIP"
-            :lock-key="`lora-clip:${item.name}`"
-            :min="LORA_STRENGTH_MIN"
-            :max="LORA_STRENGTH_MAX"
+              :value="displayStrength(item.strengthModel)"
+              :aria-label="(clipStrength ? 'Model' : 'Strength') + ' for ' + labelFor(item.name)"
+              @input="onRange(index, 'strengthModel', $event)"
+            >
+            <input
+              type="number"
+              class="lora-strength-num w-full rounded border border-[#2A2A2E] bg-[#0B0B0D] px-1 py-0.5 text-right font-mono text-[11px] text-[#F3F1EA] outline-none focus:border-[#F5A524]"
+              :min="0"
+              :max="2"
               :step="LORA_STRENGTH_STEP"
-            :fine-step="LORA_STRENGTH_STEP"
-            :coarse-step="0.25"
-            :digits="2"
-            compact
-            :disabled="disabled || !clipStrength"
-            @update:model-value="setStrength(index, 'strengthClip', $event)"
-          />
+              :disabled="disabled"
+              :value="displayStrength(item.strengthModel).toFixed(2)"
+              :aria-label="(clipStrength ? 'Model' : 'Strength') + ' value'"
+              @change="onNumber(index, 'strengthModel', $event)"
+            >
+          </label>
+          <label
+            v-if="clipStrength"
+            class="grid grid-cols-[3.25rem_minmax(0,1fr)_2.75rem] items-center gap-2 text-[11px] text-[#9A9588]"
+          >
+            <span>CLIP</span>
+            <input
+              type="range"
+              class="lora-strength-range min-w-0"
+              :min="0"
+              :max="2"
+              :step="LORA_STRENGTH_STEP"
+              :disabled="disabled"
+              :value="displayStrength(item.strengthClip)"
+              :aria-label="'CLIP for ' + labelFor(item.name)"
+              @input="onRange(index, 'strengthClip', $event)"
+            >
+            <input
+              type="number"
+              class="lora-strength-num w-full rounded border border-[#2A2A2E] bg-[#0B0B0D] px-1 py-0.5 text-right font-mono text-[11px] text-[#F3F1EA] outline-none focus:border-[#F5A524]"
+              :min="0"
+              :max="2"
+              :step="LORA_STRENGTH_STEP"
+              :disabled="disabled"
+              :value="displayStrength(item.strengthClip).toFixed(2)"
+              :aria-label="'CLIP value'"
+              @change="onNumber(index, 'strengthClip', $event)"
+            >
+          </label>
         </div>
       </li>
     </ul>
-    <p v-if="!clipStrength && model.length" class="mt-1 text-[11px] text-zinc-500">CLIP unused on this graph — model strength only.</p>
     <select
       v-if="model.length < LORA_STACK_MAX"
-      class="w-full rounded-xl border border-white/10 bg-zinc-950 outline-none ring-amber-300/40 focus:ring-2 disabled:opacity-50"
+      class="w-full rounded-lg border border-[#2A2A2E] bg-[#0B0B0D] text-[#F3F1EA] outline-none ring-[#F5A524]/40 focus:ring-2 disabled:opacity-50"
       :class="[
-        compact ? 'px-2 py-1 text-xs' : 'px-3 py-2 text-sm',
+        compact ? 'px-2 py-1.5 text-xs' : 'px-3 py-2 text-sm',
         model.length ? 'mt-2' : ''
       ]"
       :disabled="disabled || !availableNames.length"
@@ -103,13 +132,55 @@
   </div>
 </template>
 
+<style scoped>
+.lora-strength-range {
+  -webkit-appearance: none;
+  appearance: none;
+  height: 6px;
+  border-radius: 999px;
+  background: #2a2a2e;
+  outline: none;
+}
+.lora-strength-range::-webkit-slider-thumb {
+  -webkit-appearance: none;
+  appearance: none;
+  width: 14px;
+  height: 14px;
+  border-radius: 999px;
+  background: #f5a524;
+  border: 0;
+  cursor: pointer;
+}
+.lora-strength-range::-moz-range-track {
+  height: 6px;
+  border-radius: 999px;
+  background: #2a2a2e;
+}
+.lora-strength-range::-moz-range-thumb {
+  width: 14px;
+  height: 14px;
+  border-radius: 999px;
+  background: #f5a524;
+  border: 0;
+  cursor: pointer;
+}
+.lora-strength-range:disabled {
+  opacity: 0.5;
+  cursor: not-allowed;
+}
+.lora-strength-range::-webkit-slider-runnable-track {
+  height: 6px;
+  border-radius: 999px;
+  background: #2a2a2e;
+}
+</style>
+
 <script setup lang="ts">
 import {
   LORA_STACK_MAX,
-  LORA_STRENGTH_MAX,
-  LORA_STRENGTH_MIN,
+  LORA_STRENGTH_DEFAULT,
   LORA_STRENGTH_STEP,
-  clampLoraStrength,
+  loraDisplayName,
   loraFriendlyName,
   loraIdentityKey,
   loraMatchesImageEngine,
@@ -120,6 +191,9 @@ import {
   type LoraStackItem
 } from '~/utils/loras'
 
+const UI_MIN = 0
+const UI_MAX = 2
+
 const props = withDefaults(defineProps<{
   names: string[]
   filenameSubtitle?: boolean
@@ -175,6 +249,17 @@ const selectPlaceholder = computed(() => {
   return model.value.length ? props.addLabel : props.emptyLabel
 })
 
+function clampUi(raw: unknown) {
+  const value = Number(raw)
+  if (!Number.isFinite(value)) return LORA_STRENGTH_DEFAULT
+  const snapped = Math.round(value / LORA_STRENGTH_STEP) * LORA_STRENGTH_STEP
+  return Math.min(UI_MAX, Math.max(UI_MIN, Math.round(snapped * 100) / 100))
+}
+
+function displayStrength(value: unknown) {
+  return clampUi(value ?? LORA_STRENGTH_DEFAULT)
+}
+
 function onAdd(event: Event) {
   const name = (event.target as HTMLSelectElement).value
   ;(event.target as HTMLSelectElement).value = ''
@@ -186,7 +271,11 @@ function addLora(name: string) {
   if (!item || model.value.length >= LORA_STACK_MAX) return
   const taken = new Set(model.value.map(entry => loraIdentityKey(entry.name)))
   if (taken.has(loraIdentityKey(item.name))) return
-  model.value = [...model.value, item]
+  model.value = [...model.value, {
+    ...item,
+    strengthModel: clampUi(item.strengthModel ?? LORA_STRENGTH_DEFAULT),
+    strengthClip: clampUi(item.strengthClip ?? LORA_STRENGTH_DEFAULT)
+  }]
 }
 
 function removeAt(index: number) {
@@ -234,8 +323,22 @@ function onDragEnd() {
 }
 
 function setStrength(index: number, key: 'strengthModel' | 'strengthClip', value: number) {
-  const next = clampLoraStrength(value)
-  model.value = model.value.map((item, i) => i === index ? { ...item, [key]: next } : item)
+  const next = clampUi(value)
+  model.value = model.value.map((item, i) => {
+    if (i !== index) return item
+    if (!props.clipStrength) {
+      return { ...item, strengthModel: next, strengthClip: next }
+    }
+    return { ...item, [key]: next }
+  })
+}
+
+function onRange(index: number, key: 'strengthModel' | 'strengthClip', event: Event) {
+  setStrength(index, key, Number((event.target as HTMLInputElement).value))
+}
+
+function onNumber(index: number, key: 'strengthModel' | 'strengthClip', event: Event) {
+  setStrength(index, key, Number((event.target as HTMLInputElement).value))
 }
 
 function triggerFor(name: string) {
@@ -246,6 +349,10 @@ function labelFor(name: string) {
   return loraFriendlyName(name, props.aliases)
 }
 
+function stemFor(name: string) {
+  return loraDisplayName(name)
+}
+
 function saveTrigger(name: string, trigger: string) {
   emit('update-trigger', name, trigger)
 }

components/SettingsForm.vue

@@ -20,6 +20,37 @@
         :class="noticeError ? 'border-red-400/30 bg-red-950/40 text-red-100' : 'border-emerald-400/20 bg-emerald-950/30 text-emerald-100'"
       >{{ notice }}</p>
 
+      <section v-if="activePanel === 'display'" class="glass rounded-3xl p-5 sm:p-6">
+          <h2 class="font-display text-xl font-bold">Display</h2>
+          <p class="mt-1 text-sm text-zinc-400">Layout only. Jobs, engines, and library IDs stay the same.</p>
+          <div class="mt-4 grid gap-3 sm:grid-cols-2">
+            <button
+              type="button"
+              class="rounded-2xl border p-4 text-left outline-none ring-amber-300/40 focus-visible:ring-2"
+              :class="storedDisplayMode === 'classic'
+                ? 'border-amber-300/60 bg-amber-400/10'
+                : 'border-white/10 bg-zinc-950/50 hover:border-white/25'"
+              :aria-pressed="storedDisplayMode === 'classic'"
+              @click="setDisplayMode('classic')"
+            >
+              <span class="block font-display text-lg font-semibold text-zinc-50">Classic</span>
+              <span class="mt-1 block text-xs text-zinc-400">Packed header, compact selects, tight gaps.</span>
+            </button>
+            <button
+              type="button"
+              class="rounded-2xl border p-4 text-left outline-none ring-amber-300/40 focus-visible:ring-2"
+              :class="storedDisplayMode === 'studio'
+                ? 'border-amber-300/60 bg-amber-400/10'
+                : 'border-white/10 bg-zinc-950/50 hover:border-white/25'"
+              :aria-pressed="storedDisplayMode === 'studio'"
+              @click="setDisplayMode('studio')"
+            >
+              <span class="block font-display text-lg font-semibold text-zinc-50">Studio</span>
+              <span class="mt-1 block text-xs text-zinc-400">Task and model rails, locks by slots, eye toggles.</span>
+            </button>
+          </div>
+        </section>
+
       <section v-if="activePanel === 'gpu'" class="glass rounded-3xl p-5 sm:p-6">
           <h2 class="font-display text-xl font-bold">GPU & defaults</h2>
           <div class="mt-4 space-y-3 rounded-2xl border border-white/10 bg-zinc-950/50 px-4 py-3">
@@ -422,11 +453,12 @@ export interface SettingsFolder {
   unlocked: boolean
 }
 
-type SettingsPanel = 'gpu' | 'storage' | 'login' | 'folders' | 'loras' | 'keeps' | 'prompts'
+type SettingsPanel = 'display' | 'gpu' | 'storage' | 'login' | 'folders' | 'loras' | 'keeps' | 'prompts'
 type LoraKindFilter = 'all' | 'image' | 'video'
 
 const route = useRoute()
 const router = useRouter()
+const { storedDisplayMode, setDisplayMode } = useDisplayMode()
 
 const folders = ref<SettingsFolder[]>([])
 const authMode = ref('')
@@ -435,7 +467,7 @@ const notice = ref('')
 const noticeError = ref(false)
 let noticeTimer: ReturnType<typeof setTimeout> | null = null
 
-const panelIds: SettingsPanel[] = ['gpu', 'storage', 'login', 'folders', 'loras', 'keeps', 'prompts']
+const panelIds: SettingsPanel[] = ['display', 'gpu', 'storage', 'login', 'folders', 'loras', 'keeps', 'prompts']
 
 function parsePanel(raw: unknown): SettingsPanel {
   const value = String(raw || '')
@@ -456,6 +488,7 @@ const activePanel = computed({
 
 const navItems = computed(() => {
   const items: Array<{ id: SettingsPanel; label: string }> = [
+    { id: 'display', label: 'Display' },
     { id: 'gpu', label: 'GPU & defaults' },
     { id: 'storage', label: 'Storage' }
   ]

components/StudioKindCards.vue

@@ -1,7 +1,7 @@
 <template>
   <nav v-if="!advanced" aria-label="Generation type" class="grid grid-cols-3 gap-2 rounded-2xl border border-white/10 bg-zinc-950/50 p-2">
     <button v-for="kind in ['image', 'video'] as const" :key="kind" type="button" :aria-pressed="current === kind" class="rounded-xl px-3 py-3 font-semibold capitalize" :class="current === kind ? 'bg-amber-400 text-zinc-950' : 'text-zinc-300 hover:bg-white/5'" @click="emit('select', kind)">{{ kind }}</button>
-    <NuxtLink to="/music" class="rounded-xl px-3 py-3 text-center font-semibold" :aria-current="current === 'music' ? 'page' : undefined" :class="current === 'music' ? 'bg-amber-400 text-zinc-950' : 'text-zinc-300 hover:bg-white/5'">Music</NuxtLink>
+    <NuxtLink to="/music?engine=yue2" class="rounded-xl px-3 py-3 text-center font-semibold" :aria-current="current === 'music' ? 'page' : undefined" :class="current === 'music' ? 'bg-amber-400 text-zinc-950' : 'text-zinc-300 hover:bg-white/5'">Music</NuxtLink>
   </nav>
   <div v-else class="mb-4 grid grid-cols-3 gap-3">
     <div
@@ -89,7 +89,7 @@
       :class="current === 'music' ? 'border-amber-300 bg-amber-400/10 text-amber-50' : 'border-white/10 text-zinc-400'"
     >
       <NuxtLink
-        to="/music"
+        to="/music?engine=yue2"
         class="block w-full px-4 pb-3 pt-4 text-left transition"
         :class="current === 'music' ? '' : 'hover:text-zinc-200'"
       >
@@ -99,35 +99,7 @@
         </svg>
         <span class="mt-3 block font-display text-lg font-bold text-zinc-100">Music</span>
       </NuxtLink>
-      <div class="grid grid-cols-3 border-t border-white/10">
-        <button
-          type="button"
-          class="border-r border-white/10 px-3 py-2.5 text-left text-sm transition"
-          :class="engineBtnClass(current === 'music' && musicEngine === 'ace')"
-          @click="pickMusic('ace')"
-        >
-          <span class="block font-semibold">ACE</span>
-          <span class="text-xs opacity-80">Step</span>
-        </button>
-        <button
-          type="button"
-          class="border-r border-white/10 px-3 py-2.5 text-left text-sm transition"
-          :class="engineBtnClass(current === 'music' && musicEngine === 'yue')"
-          @click="pickMusic('yue')"
-        >
-          <span class="block font-semibold">YuE</span>
-          <span class="text-xs opacity-80">v1</span>
-        </button>
-        <button
-          type="button"
-          class="px-3 py-2.5 text-left text-sm transition"
-          :class="engineBtnClass(current === 'music' && musicEngine === 'yue2')"
-          @click="pickMusic('yue2')"
-        >
-          <span class="block font-semibold">YuE2</span>
-          <span class="text-xs opacity-80">3B</span>
-        </button>
-      </div>
+      <p class="border-t border-white/10 px-4 py-2 text-xs text-zinc-400">YuE2 · 3B lyrics-to-song</p>
     </div>
   </div>
 </template>
@@ -140,20 +112,19 @@ const props = withDefaults(defineProps<{
   videoEngine?: 'minimax' | 'ltx'
   videoStart?: 'still' | 'text'
   imageEngine?: 'flux' | 'krea'
-  musicEngine?: 'ace' | 'yue' | 'yue2'
+  musicEngine?: 'yue2'
 }>(), {
   ltxEnabled: false,
   videoEngine: 'minimax',
   videoStart: 'still',
   imageEngine: 'flux',
-  musicEngine: 'ace'
+  musicEngine: 'yue2'
 })
 
 const emit = defineEmits<{
   select: [kind: 'video' | 'image']
   'video-engine': [engine: 'minimax' | 'ltx']
   'image-engine': [engine: 'flux' | 'krea']
-  'music-engine': [engine: 'ace' | 'yue' | 'yue2']
 }>()
 
 function engineBtnClass(active: boolean) {
@@ -170,12 +141,4 @@ function pickImage(engine: 'flux' | 'krea') {
   emit('image-engine', engine)
   if (props.current !== 'image') emit('select', 'image')
 }
-
-function pickMusic(engine: 'ace' | 'yue' | 'yue2') {
-  if (props.current === 'music') {
-    emit('music-engine', engine)
-    return
-  }
-  void navigateTo({ path: '/music', query: { engine } })
-}
 </script>

components/studio2/Bench.vue

@@ -1,101 +1,481 @@
 <template>
- <main class="w-full min-w-0 space-y-5 p-4 md:p-6 xl:p-8">
-  <header class="flex flex-wrap items-center justify-between gap-4">
-   <div><h1 class="font-display text-3xl font-bold">Studio 2</h1><p class="text-sm text-zinc-400">One hero. Every shot.</p></div>
+ <main class="display-stack w-full min-w-0 p-4 md:p-6 xl:p-8" :data-display="displayMode">
+  <header class="flex flex-wrap items-end justify-between gap-4">
+   <div>
+    <h1 class="display-title font-display font-bold tracking-tight text-zinc-50">Studio 2</h1>
+    <p class="display-body mt-1 text-zinc-400">One hero. Every shot.</p>
+   </div>
    <div class="flex flex-wrap items-center gap-3">
-    <label class="text-sm">Folder <select v-model="form.folderId" class="rounded-lg bg-zinc-800 p-2"><option v-for="f in library.folders" :value="f.id" :key="f.id">{{ f.name }}</option></select></label>
-    <label class="text-sm"><input v-model="form.lockFace" type="checkbox"> Lock face <span class="block text-xs text-zinc-500">Face, hair, glasses, body</span></label><label class="text-sm"><input v-model="form.lockOutfit" type="checkbox"> Lock outfit <span class="block text-xs text-zinc-500">Wardrobe from hero</span></label>
-    <NuxtLink to="/library" class="text-sm text-amber-200">Library</NuxtLink><NuxtLink to="/" class="text-sm text-zinc-400">Studio 1</NuxtLink>
+    <label class="display-body text-zinc-300">Folder
+     <select v-model="form.folderId" class="display-control ml-2 border border-white/10 bg-zinc-900 px-3 py-2 outline-none ring-amber-300/40 focus:ring-2">
+      <option v-for="f in library.folders" :key="f.id" :value="f.id">{{ f.name }}</option>
+     </select>
+    </label>
+    <div v-if="isClassic && form.mode!=='music' && form.mode!=='caption'" class="flex flex-wrap items-center gap-3 text-sm text-zinc-300" aria-label="Identity locks">
+     <label class="inline-flex items-center gap-1.5"><input v-model="form.lockFace" type="checkbox" class="accent-amber-400">Lock face</label>
+     <label class="inline-flex items-center gap-1.5"><input v-model="form.lockOutfit" type="checkbox" class="accent-amber-400">Lock outfit</label>
+    </div>
+    <div v-if="isClassic && xaigen" class="flex flex-wrap items-center gap-3 text-sm text-zinc-300" aria-label="Preview hide">
+     <label class="inline-flex items-center gap-1.5"><input v-model="classicHideInput" type="checkbox" class="accent-amber-400">Hide input</label>
+     <label class="inline-flex items-center gap-1.5"><input v-model="classicHideOutput" type="checkbox" class="accent-amber-400">Hide output</label>
+    </div>
+    <nav class="flex flex-wrap items-center gap-3 text-sm" aria-label="Studio links">
+     <Studio2GpuStatus :status="gpu" />
+     <NuxtLink to="/queue" class="text-zinc-300 outline-none ring-amber-300/40 focus-visible:ring-2">Queue</NuxtLink>
+     <NuxtLink to="/library" class="text-amber-200 outline-none ring-amber-300/40 focus-visible:ring-2">Library</NuxtLink>
+     <NuxtLink to="/settings" class="text-zinc-300 outline-none ring-amber-300/40 focus-visible:ring-2">Settings</NuxtLink>
+    </nav>
    </div>
   </header>
-  <div class="flex flex-wrap justify-between gap-3 rounded-xl border border-white/10 px-4 py-2 text-sm"><Studio2GpuStatus :status="gpu" /><div v-if="xaigen" class="flex gap-4"><label><input v-model="form.hideInput" type="checkbox"> Hide input stills</label><label><input v-model="hideOutput" type="checkbox"> Hide output preview</label></div></div>
-  <nav class="flex flex-wrap gap-2"><button v-for="task in tasks" :key="task" class="rounded-full border px-4 py-2 capitalize" :class="form.mode===task?'border-amber-300 bg-amber-300 text-black':'border-white/15'" @click="chooseTask(task)">{{ task }}</button></nav>
+
+  <div class="display-stack min-w-0">
+   <label v-if="isClassic" class="display-body block text-zinc-300">Task
+    <select :value="form.mode" class="display-control mt-1 w-full max-w-xs border border-white/10 bg-zinc-900 px-3 py-2 capitalize outline-none ring-amber-300/40 focus:ring-2" @change="onTaskSelect">
+     <option v-for="task in tasks" :key="task" :value="task" class="capitalize">{{ task }}</option>
+    </select>
+   </label>
+   <nav v-else class="display-rail" aria-label="Tasks">
+    <button
+     v-for="task in tasks"
+     :key="task"
+     type="button"
+     class="display-rail__seg capitalize"
+     :aria-pressed="form.mode===task"
+     @click="chooseTask(task)"
+    >{{ task }}</button>
+   </nav>
+
+   <div v-if="form.mode!=='music' && form.mode!=='caption'" class="min-w-0">
+    <label v-if="isClassic" class="display-body block text-zinc-300">Model
+     <select v-model="form.engine" class="display-control mt-1 w-full max-w-md border border-white/10 bg-zinc-900 px-3 py-2 outline-none ring-amber-300/40 focus:ring-2">
+      <option v-for="opt in engineOptions" :key="opt.value" :value="opt.value" :disabled="opt.disabled">{{ opt.label }}{{ opt.hint ? ' — '+opt.hint : '' }}</option>
+     </select>
+    </label>
+    <div v-else class="display-rail" role="group" aria-label="Model">
+     <button
+      v-for="opt in engineOptions"
+      :key="opt.value"
+      type="button"
+      class="display-rail__seg"
+      :aria-pressed="form.engine===opt.value"
+      :aria-disabled="opt.disabled || undefined"
+      :disabled="opt.disabled"
+      :title="opt.hint || undefined"
+      @click="!opt.disabled && (form.engine=opt.value)"
+     >
+      <span class="block">{{ opt.label }}</span>
+      <span v-if="opt.hint" class="mt-0.5 block text-[11px] font-normal opacity-80">{{ opt.hint }}</span>
+     </button>
+    </div>
+   </div>
+   <p v-else-if="form.mode==='music'" class="display-body text-zinc-400">Model · YuE2</p>
+   <p v-else class="display-body text-zinc-400">Model · Qwen2.5-VL caption</p>
+  </div>
+
   <div class="grid min-w-0 gap-6 xl:grid-cols-[minmax(0,3fr)_minmax(24rem,2fr)] 2xl:gap-8">
    <section class="min-w-0 space-y-4">
-    <div class="flex min-h-80 items-center justify-center overflow-hidden rounded-3xl border border-white/10 bg-black">
-     <p v-if="selected && xaigen && (hideOutput || (selected.role==='input' && form.hideInput))" class="text-zinc-400">Output loaded · preview hidden</p>
+    <div class="display-viewer relative flex min-h-80 items-center justify-center overflow-hidden border border-white/10 bg-black">
+     <button
+      v-if="selected && xaigen && isStudio"
+      type="button"
+      class="absolute right-3 top-3 z-10 rounded-full border border-white/20 bg-zinc-950/80 p-2 text-zinc-200 outline-none ring-amber-300/50 hover:border-amber-300/50 focus-visible:ring-2"
+      :aria-pressed="!isHidden(selected)"
+      :aria-label="isHidden(selected) ? 'Show preview' : 'Hide preview'"
+      @click.stop="toggleReveal(selected.id)"
+     >
+      <svg class="h-5 w-5" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" aria-hidden="true">
+       <path v-if="isHidden(selected)" d="M3 3l18 18M10.5 10.6a2.5 2.5 0 0 0 3 3M9.4 5.1A10.4 10.4 0 0 1 12 4.5c5 0 9.3 3.1 11 7.5a11.6 11.6 0 0 1-4.2 5.1M6.2 6.2A11.5 11.5 0 0 0 1 12c1.2 2.7 3.3 4.9 5.9 6.3A10.5 10.5 0 0 0 12 19.5c1.2 0 2.4-.2 3.5-.6" stroke-linecap="round"/>
+       <template v-else>
+        <path d="M2 12s3.5-7.5 10-7.5S22 12 22 12s-3.5 7.5-10 7.5S2 12 2 12Z"/>
+        <circle cx="12" cy="12" r="3"/>
+       </template>
+      </svg>
+     </button>
+     <p v-if="selected && isHidden(selected)" class="text-zinc-400">{{ selected.kind==='track' ? 'Track loaded · preview hidden' : selected.role==='input' ? 'Still loaded · preview hidden' : 'Output hidden' }}</p>
+     <div v-else-if="selected?.kind==='track'" class="w-full max-w-xl px-6 py-4">
+      <AudioPlayer :src="`/api/library/tracks/${selected.id}/audio`" :download-name="`${selected.name || 'track'}.${selected.ext || 'wav'}`" :label="selected.name || 'Track'" />
+     </div>
      <video v-else-if="selected?.kind==='clip'" :key="selected.id" :src="`/api/library/clips/${selected.id}/video`" controls class="max-h-[65vh] w-full object-contain" />
      <img v-else-if="selected" :src="stillUrl(selected.id)" class="max-h-[65vh] max-w-full object-contain" alt="Current output">
      <p v-else class="text-zinc-500">Select a result to view it here</p>
     </div>
-    <div v-if="selected" class="flex flex-wrap items-start gap-3"><button class="rounded-full border border-white/20 px-4 py-2 text-sm" @click="restore(selected)">Use as input</button><button v-if="selected.kind==='still'" class="rounded-full border border-white/20 px-4 py-2 text-sm" @click="Object.assign(form,useAsSource(form,selected))">Use as source still</button><button v-if="selected.kind==='still'" class="rounded-full border border-white/20 px-4 py-2 text-sm" @click="Object.assign(form,useAsHero(form,selected))">Use as hero</button><button class="rounded-full border border-white/20 px-4 py-2 text-sm" @click="detailsOpen=true">Details</button><button class="rounded-full border border-red-400/40 px-4 py-2 text-sm text-red-200" @click="deleteOutput">Delete</button><VideoUpscaleControl v-if="selected.kind==='clip'" :clip-id="selected.id" queue-only @complete="refresh" /></div>
-    <div class="flex min-w-0 max-w-full gap-2 overflow-x-auto" aria-label="Current family outputs"><button v-for="o in family" :key="o.id" @click="selectOutput(o)" class="shrink-0 rounded-lg border p-1" :class="selected?.id===o.id?'border-amber-300':'border-white/10'"><span v-if="previewHidden(o,xaigen,form.hideInput,hideOutput)" class="flex h-20 w-24 items-center justify-center p-1 text-xs text-zinc-400">{{ o.role==='input'?'Still loaded · preview hidden':'Output hidden' }}</span><img v-else :src="o.kind==='clip'?`/api/library/clips/${o.id}/thumb`:stillUrl(o.id)" class="h-20 w-24 rounded object-cover" alt="Family output"></button></div>
-    <section aria-label="Source" class="rounded-2xl border border-white/10 p-3"><h2 class="mb-3 text-xs uppercase tracking-wider text-zinc-400">Source</h2><div class="flex flex-wrap gap-4">
-     <Studio2AssetSlot label="Hero" :asset="stills.find((a:any)=>a.id===form.identityStillId)" :hidden="inputHidden" upload :disabled="uploading" @pick="picker='identityStillId'" @upload="upload($event,'hero')" />
-     <Studio2AssetSlot :label="form.mode==='extend'?'Previous clip':'Start still'" :asset="sourceSlot" :hidden="inputHidden" upload :disabled="uploading" @pick="picker=form.mode==='extend'?'startClipId':'imageAId'" @upload="upload($event,'source')" />
-     <Studio2AssetSlot v-if="form.mode==='compose'" label="Second still" :asset="stills.find((a:any)=>a.id===form.imageBId)" :hidden="inputHidden" @pick="picker='imageBId'" />
-     <Studio2AssetSlot v-if="form.mode==='refine'" label="Mask" :asset="stills.find((a:any)=>a.id===form.maskId)" :hidden="inputHidden" @pick="picker='maskId'" />
-     <Studio2AssetSlot v-if="video" label="End still" :asset="stills.find((a:any)=>a.id===form.endStillId)" :hidden="inputHidden" @pick="picker='endStillId'" />
-    </div><label v-if="form.mode==='refine'" class="mt-3 block text-sm">Mask still (white = change). On-canvas mask comes later.<select v-model="form.maskId" class="mt-1 w-full rounded-xl bg-zinc-800 p-2"><option value="">Choose mask still</option><option v-for="a in stills" :key="a.id" :value="a.id">{{ assetLabel(a) }}</option></select></label><p v-if="uploading || uploadMessage" role="status" class="mt-2 text-sm text-amber-200">{{ uploading?'Uploading image…':uploadMessage }}</p></section>
-    <details v-if="video" class="rounded-xl border border-white/10 p-3"><summary>Add Guide stills</summary><div v-for="(g,i) in form.guides" :key="i" class="mt-2 flex gap-2"><Studio2AssetSlot label="Guide" :asset="stills.find((a:any)=>a.id===g.stillId)" :hidden="inputHidden" @pick="picker='guide:'+i" /><input v-model.number="g.frame" type="number" min="1" aria-label="Guide frame" class="w-20 bg-zinc-800 p-2"><button @click="form.guides.splice(i,1)">Remove</button></div><button class="mt-3 text-amber-200" @click="form.guides.push({stillId:'',frame:24})">Add guide</button></details>
-   </section>
-   <section class="min-w-0 space-y-4">
 
-    <label class="block">Prompt<textarea v-model="form.promptSections.action" rows="4" class="mt-1 w-full rounded-xl bg-zinc-900 p-3" placeholder="Describe the image or the shot." /></label>
-    <Studio2PromptControls v-if="video" v-model="form.promptSections" /><Studio2StyleChips v-else v-model="form.imageStyles" />
-    <label v-if="video" class="block">Extra constraints (optional) — do not repeat the Prompt. Example: keep the black glasses, no extra people.<textarea v-model="form.promptSections.locks" rows="2" class="mt-1 w-full rounded-xl bg-zinc-900 p-3" /></label>
-    <details v-if="video || form.mode==='edit'" class="rounded-xl border border-white/10 p-3"><summary>{{ video?'Queue more shots':'Queue extra edit passes' }}</summary><p class="mt-2 text-xs text-zinc-400">Each blank-line-separated block is another job. Hero and settings stay. Press Enter twice between shots. A wrapped sentence is still one shot.{{ video?'':' Each extra pass edits the previous result with Klein.' }}</p><textarea v-model="batchText" rows="5" class="mt-2 w-full rounded-xl bg-zinc-900 p-3" placeholder="First additional prompt&#10;&#10;Next additional prompt" /></details>
-    <button class="text-sm text-amber-200" @click="settingsOpen=true">Generation settings</button>
-    <p v-if="error" role="alert" class="text-sm text-red-300">{{ error }}</p>
-    <p v-if="form.mode==='extend'" class="text-sm text-amber-200">{{ form.startClipId ? 'Extend · '+(clips.find((c:any)=>c.id===form.startClipId)?.name || form.startClipId)+' · last frame' : 'Select the clip in the family strip or Previous clip.' }}</p>
+    <div v-if="selected" class="flex flex-wrap items-start gap-3">
+     <button v-if="selected.kind!=='track'" class="rounded-full border border-white/20 px-4 py-2 text-sm outline-none ring-amber-300/40 focus-visible:ring-2" @click="restore(selected)">Use as input</button>
+     <button v-if="selected.kind==='still'" class="rounded-full border border-white/20 px-4 py-2 text-sm outline-none ring-amber-300/40 focus-visible:ring-2" @click="Object.assign(form,useAsSource(form,selected))">Use as source still</button>
+     <button v-if="selected.kind==='still'" class="rounded-full border border-white/20 px-4 py-2 text-sm outline-none ring-amber-300/40 focus-visible:ring-2" @click="Object.assign(form,useAsHero(form,selected))">Use as hero</button>
+     <button v-if="selected.kind==='still' || (selected.kind!=='track' && selected.kind!=='clip' && selected.id)" class="rounded-full border border-amber-300/40 px-4 py-2 text-sm text-amber-100 outline-none ring-amber-300/40 focus-visible:ring-2" :disabled="submitting" @click="describeSelected">Describe</button>
+     <button v-if="selected.kind!=='track'" class="rounded-full border border-white/20 px-4 py-2 text-sm outline-none ring-amber-300/40 focus-visible:ring-2" @click="detailsOpen=true">Details</button>
+     <button class="rounded-full border border-red-400/40 px-4 py-2 text-sm text-red-200 outline-none ring-amber-300/40 focus-visible:ring-2" @click="deleteOutput">Delete</button>
+     <VideoUpscaleControl v-if="selected.kind==='clip'" :clip-id="selected.id" queue-only @complete="refresh" />
+    </div>
+
+    <div class="flex min-w-0 max-w-full gap-2 overflow-x-auto" aria-label="Current family outputs">
+     <div
+      v-for="o in family"
+      :key="o.id"
+      class="relative shrink-0 rounded-lg border p-1"
+      :class="selected?.id===o.id ? 'border-amber-300' : 'border-white/10'"
+     >
+      <button type="button" class="block outline-none ring-amber-300/40 focus-visible:ring-2" @click="selectOutput(o)">
+       <span v-if="isHidden(o)" class="flex h-20 w-24 items-center justify-center p-1 text-xs text-zinc-400">{{ o.kind==='track' ? 'Audio hidden' : o.role==='input' ? 'Still loaded · preview hidden' : 'Output hidden' }}</span>
+       <span v-else-if="o.kind==='track'" class="flex h-20 w-24 items-center justify-center rounded bg-zinc-900 text-xs text-amber-100">Audio</span>
+       <img v-else :src="o.kind==='clip' ? `/api/library/clips/${o.id}/thumb` : stillUrl(o.id)" class="h-20 w-24 rounded object-cover" alt="Family output">
+      </button>
+      <button
+       v-if="xaigen && isStudio"
+       type="button"
+       class="absolute right-1 top-1 rounded-md border border-white/20 bg-zinc-950/80 p-1 text-zinc-200 outline-none ring-amber-300/50 focus-visible:ring-2"
+       :aria-label="isHidden(o) ? 'Show tile' : 'Hide tile'"
+       :aria-pressed="!isHidden(o)"
+       @click.stop="toggleReveal(o.id)"
+      >
+       <svg class="h-3.5 w-3.5" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.75" aria-hidden="true">
+        <path v-if="isHidden(o)" d="M3 3l18 18M10.5 10.6a2.5 2.5 0 0 0 3 3M9.4 5.1A10.4 10.4 0 0 1 12 4.5c5 0 9.3 3.1 11 7.5a11.6 11.6 0 0 1-4.2 5.1M6.2 6.2A11.5 11.5 0 0 0 1 12c1.2 2.7 3.3 4.9 5.9 6.3A10.5 10.5 0 0 0 12 19.5c1.2 0 2.4-.2 3.5-.6" stroke-linecap="round"/>
+        <template v-else>
+         <path d="M2 12s3.5-7.5 10-7.5S22 12 22 12s-3.5 7.5-10 7.5S2 12 2 12Z"/>
+         <circle cx="12" cy="12" r="3"/>
+        </template>
+       </svg>
+      </button>
+     </div>
+    </div>
+
+    <section v-if="form.mode!=='music' && form.mode!=='caption'" aria-label="Source" class="display-card">
+     <div class="mb-3 flex flex-wrap items-start justify-between gap-3">
+      <h2 class="text-xs uppercase tracking-wider text-zinc-400">Source</h2>
+      <div v-if="isStudio" class="flex flex-wrap gap-4 text-sm text-zinc-300" aria-label="Identity locks">
+       <template v-if="form.engine!=='qwen21'">
+       <label class="flex items-start gap-2"><input v-model="form.lockFace" type="checkbox" class="mt-1 accent-amber-400"><span>Lock face<span class="block text-xs text-zinc-500">Hero encode + lock on Klein / Krea stills</span></span></label>
+       <label class="flex items-start gap-2"><input v-model="form.lockOutfit" type="checkbox" class="mt-1 accent-amber-400"><span>Lock outfit<span class="block text-xs text-zinc-500">Hero encode + lock on Klein / Krea stills</span></span></label>
+       </template>
+      </div>
+     </div>
+     <div class="flex flex-wrap gap-4">
+      <Studio2AssetSlot label="Hero" :asset="stills.find((a:any)=>a.id===form.identityStillId)" :hidden="slotHidden(form.identityStillId)" upload :disabled="uploading" @pick="picker='identityStillId'" @upload="upload($event,'hero')" />
+      <Studio2AssetSlot :label="form.mode==='extend'?'Previous clip':'Start still'" :asset="sourceSlot" :hidden="slotHidden(form.mode==='extend'?form.startClipId:form.imageAId)" upload :disabled="uploading" @pick="picker=form.mode==='extend'?'startClipId':'imageAId'" @upload="upload($event,'source')" />
+      <Studio2AssetSlot v-if="form.mode==='compose'" label="Second still" :asset="stills.find((a:any)=>a.id===form.imageBId)" :hidden="slotHidden(form.imageBId)" @pick="picker='imageBId'" />
+      <Studio2AssetSlot v-if="form.mode==='refine'" label="Mask" :asset="stills.find((a:any)=>a.id===form.maskId)" :hidden="slotHidden(form.maskId)" @pick="picker='maskId'" />
+      <Studio2AssetSlot v-if="video" label="End still" :asset="stills.find((a:any)=>a.id===form.endStillId)" :hidden="slotHidden(form.endStillId)" @pick="picker='endStillId'" />
+     </div>
+     <label v-if="form.mode==='refine'" class="mt-3 block text-sm">Mask still (white = change). On-canvas mask comes later.
+      <select v-model="form.maskId" class="mt-1 w-full rounded-xl bg-zinc-800 p-2"><option value="">Choose mask still</option><option v-for="a in stills" :key="a.id" :value="a.id">{{ assetLabel(a) }}</option></select>
+     </label>
+     <p v-if="uploading || uploadMessage" role="status" class="mt-2 text-sm text-amber-200">{{ uploading?'Uploading image…':uploadMessage }}</p>
+    </section>
+
+    <details v-if="video" class="rounded-xl border border-white/10 p-3"><summary>Add Guide stills</summary>
+     <div v-for="(g,i) in form.guides" :key="i" class="mt-2 flex gap-2">
+      <Studio2AssetSlot label="Guide" :asset="stills.find((a:any)=>a.id===g.stillId)" :hidden="slotHidden(g.stillId)" @pick="picker='guide:'+i" />
+      <input v-model.number="g.frame" type="number" min="1" aria-label="Guide frame" class="w-20 bg-zinc-800 p-2">
+      <button type="button" @click="form.guides.splice(i,1)">Remove</button>
+     </div>
+     <button type="button" class="mt-3 text-amber-200" @click="form.guides.push({stillId:'',frame:24})">Add guide</button>
+    </details>
+   </section>
+
+   <section class="min-w-0 space-y-4">
+    <template v-if="form.mode==='music'">
+     <p class="text-sm text-zinc-400">YuE2 · lyrics-to-song · plan → synthesize → decode</p>
+     <label class="block text-sm">Genre / style tags
+      <textarea v-model="music.tags" rows="3" class="mt-1 w-full rounded-xl bg-zinc-900 p-3 outline-none ring-amber-300/40 focus:ring-2" :placeholder="tagsPlaceholder" />
+     </label>
+     <label class="block text-sm">Lyrics
+      <textarea v-model="music.lyrics" rows="8" class="mt-1 w-full rounded-xl bg-zinc-900 p-3 outline-none ring-amber-300/40 focus:ring-2" :placeholder="lyricsPlaceholder" />
+     </label>
+     <div class="grid grid-cols-2 gap-3">
+      <label class="block text-sm">Target length (sec)
+       <input v-model.number="music.duration" type="number" min="15" max="150" class="mt-1 w-full rounded-xl bg-zinc-900 p-3 outline-none ring-amber-300/40 focus:ring-2">
+       <span class="mt-1 block text-[11px] text-zinc-500">Caps semantic tokens (~25/s). Actual length follows the model end token — not an FFmpeg trim.</span>
+      </label>
+      <label class="block text-sm">Seed
+       <input v-model="music.seed" placeholder="random" class="mt-1 w-full rounded-xl bg-zinc-900 p-3 outline-none ring-amber-300/40 focus:ring-2">
+      </label>
+     </div>
+     <label class="block text-sm">Track name (optional)
+      <input v-model="music.name" class="mt-1 w-full rounded-xl bg-zinc-900 p-3 outline-none ring-amber-300/40 focus:ring-2" placeholder="Optional track name">
+     </label>
+    </template>
+    <template v-else-if="form.mode==='caption'">
+     <p class="text-sm text-zinc-400">Qwen2.5-VL · image-to-text · exclusive GPU · unloads after one shot</p>
+     <div class="display-card space-y-3 p-3">
+      <Studio2AssetSlot label="Still to describe" :asset="captionStill" :hidden="slotHidden(captionStillId)" upload :disabled="uploading" @pick="picker='captionStillId'" @upload="uploadCaption($event)" />
+      <label class="block text-sm text-zinc-300">Caption style
+       <select v-model="captionStyle" class="mt-1 w-full rounded-xl border border-white/10 bg-zinc-900 p-3 outline-none ring-amber-300/40 focus:ring-2">
+        <option v-for="style in captionStyles" :key="style" :value="style">{{ style }}</option>
+       </select>
+      </label>
+      <p v-if="captionStyle==='delta'" class="text-xs text-zinc-500">Delta describes only changes from the locked look (bob / pink tank / blue shorts / Slytherin robe / barefoot).</p>
+     </div>
+    </template>
+    <template v-else>
+     <label class="block">Prompt<textarea v-model="form.promptSections.action" rows="4" class="mt-1 w-full rounded-xl bg-zinc-900 p-3 outline-none ring-amber-300/40 focus:ring-2" placeholder="Describe the image or the shot." /></label>
+     <Studio2PromptControls v-if="video" v-model="form.promptSections" />
+     <Studio2StyleChips v-else v-model="form.imageStyles" />
+     <label v-if="video" class="block">Extra constraints (optional) — do not repeat the Prompt. Example: keep the black glasses, no extra people.
+      <textarea v-model="form.promptSections.locks" rows="2" class="mt-1 w-full rounded-xl bg-zinc-900 p-3 outline-none ring-amber-300/40 focus:ring-2" />
+     </label>
+     <details v-if="video || form.mode==='edit'" class="rounded-xl border border-white/10 p-3">
+      <summary>{{ video?'Queue more shots':'Queue extra edit passes' }}</summary>
+      <p class="mt-2 text-xs text-zinc-400">Each blank-line-separated block is another job. Hero and settings stay. Press Enter twice between shots. A wrapped sentence is still one shot.{{ video?'':' Each extra pass edits the previous result with Klein.' }}</p>
+      <textarea v-model="batchText" rows="5" class="mt-2 w-full rounded-xl bg-zinc-900 p-3 outline-none ring-amber-300/40 focus:ring-2" placeholder="First additional prompt&#10;&#10;Next additional prompt" />
+     </details>
+     <button type="button" class="text-sm text-amber-200 outline-none ring-amber-300/40 focus-visible:ring-2" @click="settingsOpen=true">Generation settings</button>
      <p v-if="!video && form.engine==='krea'" class="text-sm text-zinc-400">Krea does not attach the hero identity reference.</p>
-    <button :disabled="submitting" @click="queue" class="w-full rounded-2xl bg-amber-300 p-4 font-semibold text-black disabled:opacity-50">{{ submitting?'Queuing…':'Queue' }}</button>
+     <p v-if="!video && form.engine==='qwen21'" class="text-sm text-zinc-400">Qwen 2.1 is T2I — no hero, locks, or LoRA stack in this build.</p>
+     <p v-if="form.mode==='extend'" class="text-sm text-amber-200">{{ form.startClipId ? 'Extend · '+(clips.find((c:any)=>c.id===form.startClipId)?.name || form.startClipId)+' · last frame' : 'Select the clip in the family strip or Previous clip.' }}</p>
+    </template>
+
+    <p v-if="error" role="alert" class="text-sm text-red-300">{{ error }}</p>
+    <button type="button" :disabled="submitting" class="display-cta w-full p-4 outline-none disabled:opacity-50" @click="queue">{{ submitting?'Queuing…':'Queue' }}</button>
+
     <h2 class="text-sm font-semibold text-zinc-300">Job queue</h2>
-    <div v-for="job in upscales.filter(j=>j.folderId===form.folderId && ['queued','running'].includes(j.status))" :key="job.id" class="space-y-2 rounded-xl border border-amber-300/25 bg-zinc-900 p-4 text-sm" role="status"><div class="flex justify-between gap-3"><strong>Video upscale</strong><span class="capitalize text-amber-200">{{ job.status }}</span></div><p class="truncate text-zinc-300">{{ job.name }}</p><p class="text-xs text-zinc-400">{{ job.width?job.width+' × '+job.height+' · ':'' }}{{ Math.round((job.elapsedMs||0)/1000) }}s elapsed</p><progress class="h-2 w-full accent-amber-300" max="100" :value="job.progress || 0" /><p class="text-xs text-zinc-400">{{ job.message }}</p></div>
+    <div v-for="job in upscales.filter(j=>j.folderId===form.folderId && ['queued','running'].includes(j.status))" :key="job.id" class="space-y-2 rounded-xl border border-amber-300/25 bg-zinc-900 p-4 text-sm" role="status">
+     <div class="flex justify-between gap-3"><strong>Video upscale</strong><span class="capitalize text-amber-200">{{ job.status }}</span></div>
+     <p class="truncate text-zinc-300">{{ job.name }}</p>
+     <p class="text-xs text-zinc-400">{{ job.width?job.width+' × '+job.height+' · ':'' }}{{ Math.round((job.elapsedMs||0)/1000) }}s elapsed</p>
+     <progress class="h-2 w-full accent-amber-300" max="100" :value="job.progress || 0" />
+     <p class="text-xs text-zinc-400">{{ job.message }}</p>
+    </div>
+    <div v-for="job in captionJobs" :key="job.id" class="space-y-2 rounded-xl border border-amber-300/25 bg-zinc-900 p-3 text-sm" aria-live="polite" role="status">
+     <div class="flex justify-between gap-3">
+      <div class="flex min-w-0 items-start gap-3">
+       <img v-if="job.request?.stillId || job.request?.imageAId" :src="stillUrl(job.request.stillId || job.request.imageAId)" class="h-14 w-14 shrink-0 rounded object-cover" alt="Caption source">
+       <div class="min-w-0">
+        <strong class="block truncate">Describe · {{ job.request?.captionStyle || 'caption' }}</strong>
+        <span class="capitalize text-amber-200">{{ job.state==='waiting'?'queued':job.state }}</span>
+       </div>
+      </div>
+      <button v-if="!['complete','failed','cancelled'].includes(job.state)" type="button" class="shrink-0 text-zinc-300" @click="cancel(job)">Cancel</button>
+     </div>
+     <p class="text-xs text-zinc-400">{{ elapsedCaption(job) }}</p>
+     <p v-if="job.resultText" class="whitespace-pre-wrap text-zinc-200">{{ job.resultText }}</p>
+     <div v-if="job.resultText" class="flex flex-wrap gap-2">
+      <button type="button" class="rounded-full border border-white/20 px-3 py-1.5 text-xs outline-none ring-amber-300/40 focus-visible:ring-2" @click="copyCaption(job.resultText)">Copy</button>
+      <button type="button" class="rounded-full border border-amber-300/40 px-3 py-1.5 text-xs text-amber-100 outline-none ring-amber-300/40 focus-visible:ring-2" @click="useCaptionAsPrompt(job.resultText)">Use as prompt</button>
+     </div>
+     <p v-if="job.error" role="alert" class="text-red-300">{{ job.error }}</p>
+    </div>
+    <div v-for="job in musicJobs" :key="job.id" class="space-y-2 rounded-xl border border-amber-300/25 bg-zinc-900 p-3 text-sm" aria-live="polite" role="status">
+     <div class="flex justify-between gap-3">
+      <strong class="truncate">{{ job.name || 'YuE2' }}</strong>
+      <span class="capitalize text-amber-200">{{ job.live?.status || job.status }}</span>
+     </div>
+     <p class="text-xs text-zinc-400">YuE2 · target {{ job.duration || '—' }}s<span v-if="job.live?.message"> · {{ job.live.message }}</span></p>
+     <progress v-if="job.live?.progress" class="h-2 w-full accent-amber-300" max="100" :value="job.live.progress || 0" />
+     <p v-if="job.lastError || job.live?.error" role="alert" class="text-red-300">{{ job.lastError || job.live?.error }}</p>
+    </div>
     <Studio2JobProgress v-for="job in benchJobs" :key="job.id" :job="job" :now="now" @select="focusJob(job)" @cancel="cancel(job)" />
-    <button v-if="folderJobs.some(j=>['complete','failed','cancelled'].includes(j.state))" class="text-sm text-zinc-400" @click="clearFinished">Clear finished</button>
+    <button v-if="folderJobs.some(j=>['complete','failed','cancelled'].includes(j.state)) || musicJobs.some(j=>['complete','done','error','cancelled'].includes(j.status)) || captionJobs.some(j=>['complete','failed','cancelled'].includes(j.state))" type="button" class="text-sm text-zinc-400 outline-none ring-amber-300/40 focus-visible:ring-2" @click="clearFinished">Clear finished</button>
    </section>
   </div>
+
   <div v-if="settingsOpen || detailsOpen" class="fixed inset-0 z-50 bg-black/70" @click.self="settingsOpen=false;detailsOpen=false">
-   <aside class="absolute inset-y-0 right-0 w-full max-w-md overflow-auto border-l border-white/15 bg-zinc-950 p-6"><button class="float-right" @click="settingsOpen=false;detailsOpen=false">Close</button>
-    <template v-if="settingsOpen"><h2 class="mb-6 text-xl font-semibold">Generation settings</h2><label class="mb-4 block">Engine<select v-model="form.engine" class="mt-1 w-full rounded bg-zinc-800 p-2"><template v-if="video"><option value="minimax">MiniMax H3</option></template><template v-else><option value="flux">Flux Klein 9B Base</option><option value="krea" :disabled="['edit','compose','iterate'].includes(form.mode)">Krea 2 (Generate / Refine)</option></template></select></label><AspectRatioPick v-model="form.settings.aspect" heading="Aspect" :options="aspectOptions" /><p class="my-3 text-sm text-zinc-400">{{ resolvedSize.width }} × {{ resolvedSize.height }} · fit with padding, never stretch</p><label v-for="key in ['steps','cfg',...(video?['duration','fps']:[])]" :key="key" class="mb-4 block capitalize">{{ key }}<input v-model.number="form.settings[key]" type="number" class="mt-1 w-full rounded bg-zinc-800 p-2"></label><label class="mb-4 block">Seed · {{ form.settings.seedMode==='locked'?'Locked':'Random each job' }}<input :value="form.settings.seed" type="number" min="0" max="2147483647" class="mt-1 w-full rounded bg-zinc-800 p-2" @input="form.settings.seed=Number(($event.target as HTMLInputElement).value);form.settings.seedMode='locked'"></label><button class="mb-4 text-sm text-amber-200" @click="form.settings.seedMode='random';form.settings.seed=randomSeed()">Random</button><label v-if="video"><input v-model="form.settings.turbo" type="checkbox"> Turbo</label><LoraSelect v-model="form.settings.loraStack" class="mt-5" :names="pickerLoras" :use="currentLoraUse" :clip-strength="true" filename-subtitle :aliases="{...loraNames.aliases,...library.loraAliases}" :triggers="{...loraNames.triggers,...library.loraTriggers}" missing-label="No cached LoRAs yet — Refresh to check Comfy" @insert-trigger="form.promptSections.action += ' '+$event" /><button class="mt-2 text-sm text-amber-200" @click="loadLoras(true)">Refresh LoRAs</button><details class="mt-6"><summary>Compiled prompt</summary><pre class="whitespace-pre-wrap text-xs text-zinc-400">{{ compiled }}</pre><p v-if="!video" class="mt-3 text-xs text-zinc-400">Negative: {{ stylePrompt(form.imageStyles,true) || 'None' }}</p></details></template>
-    <template v-else><h2 class="mb-6 text-xl font-semibold">Output details</h2><Studio2Details :metadata="selected?.studio2" /></template>
+   <aside class="absolute inset-y-0 right-0 w-full max-w-md overflow-auto border-l border-white/15 bg-zinc-950 p-6">
+    <button type="button" class="float-right outline-none ring-amber-300/40 focus-visible:ring-2" @click="settingsOpen=false;detailsOpen=false">Close</button>
+    <template v-if="settingsOpen">
+     <h2 class="mb-6 text-xl font-semibold">Generation settings</h2>
+     <AspectRatioPick v-model="form.settings.aspect" heading="Aspect" :options="aspectOptions" />
+     <p class="my-3 text-sm text-zinc-400">{{ resolvedSize.width }} × {{ resolvedSize.height }} · fit with padding, never stretch</p>
+     <label v-for="key in ['steps','cfg',...(video?['duration','fps']:[])]" :key="key" class="mb-4 block capitalize">{{ key }}
+      <input v-model.number="form.settings[key]" type="number" class="mt-1 w-full rounded bg-zinc-800 p-2 outline-none ring-amber-300/40 focus:ring-2">
+     </label>
+     <label class="mb-4 block">Seed · {{ form.settings.seedMode==='locked'?'Locked':'Random each job' }}
+      <input :value="form.settings.seed" type="number" min="0" max="2147483647" class="mt-1 w-full rounded bg-zinc-800 p-2 outline-none ring-amber-300/40 focus:ring-2" @input="form.settings.seed=Number(($event.target as HTMLInputElement).value);form.settings.seedMode='locked'">
+     </label>
+     <button type="button" class="mb-4 text-sm text-amber-200 outline-none ring-amber-300/40 focus-visible:ring-2" @click="form.settings.seedMode='random';form.settings.seed=randomSeed()">Random</button>
+     <label v-if="video"><input v-model="form.settings.turbo" type="checkbox" class="accent-amber-400"> Turbo</label>
+     <template v-if="form.engine!=='qwen21'">
+     <LoraSelect v-model="form.settings.loraStack" class="mt-5" :names="pickerLoras" :use="currentLoraUse" :clip-strength="!video" filename-subtitle :aliases="{...loraNames.aliases,...library.loraAliases}" :triggers="{...loraNames.triggers,...library.loraTriggers}" missing-label="No cached LoRAs yet — Refresh to check Comfy" @insert-trigger="form.promptSections.action += ' '+$event" />
+     <button type="button" class="mt-2 text-sm text-amber-200 outline-none ring-amber-300/40 focus-visible:ring-2" @click="loadLoras(true)">Refresh LoRAs</button>
+     </template>
+     <details class="mt-6"><summary>Compiled prompt</summary>
+      <pre class="whitespace-pre-wrap text-xs text-zinc-400">{{ compiled }}</pre>
+      <p v-if="!video" class="mt-3 text-xs text-zinc-400">Negative: {{ stylePrompt(form.imageStyles,true) || 'None' }}</p>
+     </details>
+    </template>
+    <template v-else>
+     <h2 class="mb-6 text-xl font-semibold">Output details</h2>
+     <Studio2Details :metadata="selected?.studio2" />
+    </template>
    </aside>
   </div>
+
   <div v-if="picker" class="fixed inset-0 z-50 flex items-center justify-center bg-black/80 p-6" @click.self="picker=''">
-   <section role="dialog" aria-modal="true" aria-label="Choose source asset" class="max-h-[80vh] w-full max-w-4xl overflow-auto rounded-2xl border border-white/15 bg-zinc-950 p-5"><div class="mb-4 flex justify-between"><h2>Choose {{ picker==='identityStillId'?'hero':'source' }}</h2><button @click="picker=''">Close</button></div><button class="mb-4 text-sm text-zinc-400" @click="pickAsset(null)">Clear slot</button><div class="grid grid-cols-2 gap-3 sm:grid-cols-4 lg:grid-cols-5"><button v-for="a in pickerAssets" :key="a.id" class="rounded-xl border border-white/15 p-2 text-left" @click="pickAsset(a)"><span v-if="inputHidden" class="flex h-24 items-center text-xs text-zinc-400">Still loaded · preview hidden</span><img v-else :src="a.kind==='clip'?'/api/library/clips/'+a.id+'/thumb':stillUrl(a.id)" class="h-24 w-full object-contain" alt="Source option"><span class="mt-2 block truncate text-xs">{{ assetLabel(a) }}</span></button></div></section>
+   <section role="dialog" aria-modal="true" aria-label="Choose source asset" class="max-h-[80vh] w-full max-w-4xl overflow-auto rounded-2xl border border-white/15 bg-zinc-950 p-5">
+    <div class="mb-4 flex justify-between"><h2>Choose {{ picker==='identityStillId'?'hero':'source' }}</h2><button type="button" @click="picker=''">Close</button></div>
+    <button type="button" class="mb-4 text-sm text-zinc-400" @click="pickAsset(null)">Clear slot</button>
+    <div class="grid grid-cols-2 gap-3 sm:grid-cols-4 lg:grid-cols-5">
+     <button v-for="a in pickerAssets" :key="a.id" type="button" class="rounded-xl border border-white/15 p-2 text-left outline-none ring-amber-300/40 focus-visible:ring-2" @click="pickAsset(a)">
+      <span v-if="isHidden(a)" class="flex h-24 items-center text-xs text-zinc-400">Still loaded · preview hidden</span>
+      <img v-else :src="a.kind==='clip'?'/api/library/clips/'+a.id+'/thumb':stillUrl(a.id)" class="h-24 w-full object-contain" alt="Source option">
+      <span class="mt-2 block truncate text-xs">{{ assetLabel(a) }}</span>
+     </button>
+    </div>
+   </section>
   </div>
  </main>
 </template>
+
 <script setup lang="ts">
 import {createLoraMemory,loraUse} from '~/shared/studio2/lora-memory.mjs'
 import {filterLorasForStudio,filterLorasForUse} from '~/utils/loras'
 import {compilePrompt,sectionsFrom} from '~/shared/studio2/contracts.mjs'
-import {aspectOptions,resolveSize,bindClip,restoreAsset,extendSelected,changeSource,restoreAll,useAsHero,useAsSource} from '~/shared/studio2/input.mjs'
-import {folderAssets,newestOutput,assetLabel,reconcileSelection,activeJobs,previewHidden} from '~/shared/studio2/viewer.mjs'
-import {queueJobs} from '~/shared/studio2/queue-view.mjs'
+import {CAPTION_STYLES} from '~/shared/studio2/caption.mjs'
+import {aspectOptions,resolveSize,bindClip,changeSource,restoreAll,useAsHero,useAsSource} from '~/shared/studio2/input.mjs'
+import {folderAssets,newestOutput,assetLabel,reconcileSelection,previewHidden} from '~/shared/studio2/viewer.mjs'
+import {queueJobs,elapsedSeconds} from '~/shared/studio2/queue-view.mjs'
 import {gpuStatus} from '~/shared/studio2/status.mjs'
 import {applyUploadedStill} from '~/shared/studio2/upload.mjs'
-const xaigen=useRuntimeConfig().public.studio==='xaigen',hideOutput=ref(false),picker=ref('')
-const inputHidden=computed(()=>xaigen && form.hideInput)
-const uploading=ref(false),uploadMessage=ref('')
 import {parseBatch,stylePrompt} from '~/shared/studio2/styles.mjs'
 import {randomSeed} from '~/shared/studio2/seed.mjs'
+import {DEFAULT_MUSIC_LYRICS,DEFAULT_MUSIC_TAGS} from '~/utils/music'
+
+const xaigen=useRuntimeConfig().public.studio==='xaigen'
+const {displayMode,isStudio,isClassic}=useDisplayMode()
+const classicHideInput=ref(true)
+const classicHideOutput=ref(true)
+const picker=ref('')
+const uploading=ref(false),uploadMessage=ref('')
+const revealed=ref<Set<string>>(new Set())
 const measured=reactive<Record<string,{width:number;height:number}>>({}),seenOutputs=new Set<string>()
 const loraNames=ref<any>({image:[],video:[]}),now=ref(Date.now()),healthState=ref<any>(null),upscales=ref<any[]>([])
-const gpu=computed(()=>gpuStatus(jobs.value,healthState.value,upscales.value))
-const library=ref<any>({folders:[],stills:[],clips:[]}),jobs=ref<any[]>([]),selected=ref<any>(null)
-const form=reactive<any>({mode:'generate',engine:'flux',imageStyles:{positive:[],negative:[]},lockFace:true,lockOutfit:true,hideInput:false,folderId:'',identityStillId:'',imageAId:'',imageBId:'',maskId:'',startClipId:'',endStillId:'',guides:[],promptSections:sectionsFrom({}),settings:{aspect:'auto',fps:24,steps:24,cfg:4,seed:null,seedMode:'random',duration:5,turbo:true,loraStack:[]}})
+const library=ref<any>({folders:[],stills:[],clips:[],tracks:[]}),jobs=ref<any[]>([]),selected=ref<any>(null)
+const studioQueue=ref<any[]>([])
+const liveMusic=ref<Record<string,any>>({})
+const form=reactive<any>({mode:'generate',engine:'flux',imageStyles:{positive:[],negative:[]},lockFace:true,lockOutfit:true,hideInput:xaigen,folderId:'',identityStillId:'',imageAId:'',imageBId:'',maskId:'',startClipId:'',endStillId:'',guides:[],promptSections:sectionsFrom({}),settings:{aspect:'auto',fps:24,steps:24,cfg:4,seed:null,seedMode:'random',duration:5,turbo:true,loraStack:[]}})
+const music=reactive({tags:DEFAULT_MUSIC_TAGS,lyrics:DEFAULT_MUSIC_LYRICS,duration:60,seed:'',name:''})
+const tagsPlaceholder=DEFAULT_MUSIC_TAGS,lyricsPlaceholder=DEFAULT_MUSIC_LYRICS
+const captionStyles=CAPTION_STYLES
+const captionStyle=ref('descriptive')
+const captionStillId=ref('')
 let loraMemory=createLoraMemory(xaigen)
 const currentLoraUse=computed(()=>loraUse(form))
 const pickerLoras=computed(()=>!xaigen && currentLoraUse.value==='ltx'?[]:filterLorasForUse(filterLorasForStudio([...loraNames.value.image,...loraNames.value.video],xaigen),currentLoraUse.value))
 watch(()=>[currentLoraUse.value,form.settings.loraStack],()=>loraMemory.sync(form),{deep:true,flush:'sync'})
 const route=useRoute()
-const tasks=['generate','edit','compose','refine','video','extend'],settingsOpen=ref(false),detailsOpen=ref(false),submitting=ref(false),error=ref(''),batchText=ref('')
-const video=computed(()=>['video','extend'].includes(form.mode)),stills=computed(()=>library.value.stills.filter((s:any)=>s.folderId===form.folderId)),clips=computed(()=>library.value.clips.filter((s:any)=>s.folderId===form.folderId)),compiled=computed(()=>compilePrompt(form.promptSections,video.value,form))
+const tasks=['generate','edit','compose','refine','video','extend','music','caption']
+const settingsOpen=ref(false),detailsOpen=ref(false),submitting=ref(false),error=ref(''),batchText=ref('')
+const video=computed(()=>['video','extend'].includes(form.mode))
+const kreaSticky=computed(()=>['edit','compose','iterate'].includes(form.mode))
+const qwenOnlyGenerate=computed(()=>form.mode!=='generate')
+const engineOptions=computed(()=>{
+ if(video.value){
+  return [
+   {value:'minimax',label:'MiniMax H3',disabled:false,hint:''},
+   {value:'ltx',label:'LTX',disabled:!xaigen,hint:xaigen?'':'xAIGen only'}
+  ]
+ }
+ return [
+  {value:'flux',label:'Klein',disabled:false,hint:''},
+  {value:'krea',label:'Krea 2',disabled:kreaSticky.value,hint:kreaSticky.value?'sticky hero = Klein':''},
+  {value:'qwen21',label:'Qwen 2.1',disabled:qwenOnlyGenerate.value,hint:qwenOnlyGenerate.value?'Qwen 2.1 is T2I in this build':''}
+ ]
+})
+watch(engineOptions,opts=>{
+ const current=opts.find(o=>o.value===form.engine)
+ if(!current || current.disabled)form.engine=opts.find(o=>!o.disabled)?.value || form.engine
+})
+watch(()=>form.engine,(engine,prev)=>{
+ if(engine==='qwen21' && prev!=='qwen21'){
+  form.settings.steps=25
+  form.settings.cfg=1
+  form.lockFace=false
+  form.lockOutfit=false
+  form.settings.loraStack=[]
+ }
+})
+const stills=computed(()=>library.value.stills.filter((s:any)=>s.folderId===form.folderId))
+const clips=computed(()=>library.value.clips.filter((s:any)=>s.folderId===form.folderId))
+const compiled=computed(()=>compilePrompt(form.promptSections,video.value,form))
 const stillUrl=(id:string)=>`/api/library/stills/${id}`
 const folderJobs=computed(()=>jobs.value.filter(j=>j.request.folderId===form.folderId))
-const benchJobs=computed(()=>queueJobs(folderJobs.value))
+const captionJobs=computed(()=>folderJobs.value.filter(j=>j.request?.mode==='caption'))
+const benchJobs=computed(()=>queueJobs(folderJobs.value.filter(j=>j.request?.mode!=='caption')))
+const musicJobs=computed(()=>studioQueue.value
+ .filter((j:any)=>j.kind==='music' && j.folderId===form.folderId && !['complete','done'].includes(j.status))
+ .map((j:any)=>({...j,live:j.liveJobId?liveMusic.value[j.liveJobId]:null})))
+const gpu=computed(()=>gpuStatus(jobs.value,healthState.value,upscales.value,musicJobs.value,captionJobs.value))
+const captionStill=computed(()=>stills.value.find((s:any)=>s.id===captionStillId.value))
 const sourceSlot=computed(()=>{const clip=clips.value.find((a:any)=>a.id===form.startClipId);return form.mode==='extend'?(clip?{...clip,kind:'clip'}:null):inputStill.value})
 const pickerAssets=computed(()=>picker.value==='startClipId'?clips.value.map((a:any)=>({...a,kind:'clip'})):stills.value.map((a:any)=>({...a,kind:'still'})))
-const activeJob=computed(()=>folderJobs.value.find(j=>!['complete','failed','cancelled'].includes(j.state)))
 const inputStill=computed(()=>stills.value.find((s:any)=>s.id===form.imageAId))
 const family=computed(()=>folderAssets(library.value,jobs.value,form.folderId))
 const sourceAsset=computed(()=>form.startClipId?clips.value.find((c:any)=>c.id===form.startClipId):stills.value.find((s:any)=>s.id===(form.imageAId||form.identityStillId)))
 const sourceSize=computed(()=>measured[sourceAsset.value?.id] || sourceAsset.value)
-const resolvedSize=computed(()=>resolveSize(form.settings.aspect,sourceSize.value))
+const resolvedSize=computed(()=>resolveSize(form.settings.aspect,sourceSize.value,undefined,form.engine))
+
+function elapsedCaption(job:any){
+ const seconds=elapsedSeconds(job,now.value)
+ return `${Math.floor(seconds/60)}m ${seconds%60}s ${job.startedAt||job.state==='rendering'||job.state==='saving'?'elapsed':'waiting'}`
+}
+async function copyCaption(text:string){
+ try{await navigator.clipboard.writeText(text)}catch{error.value='Could not copy caption.'}
+}
+function useCaptionAsPrompt(text:string){
+ const current=String(form.promptSections.action||'').trim()
+ form.promptSections.action=current?`${current}\n\n${text}`:text
+ if(form.mode==='caption' || form.mode==='music')form.mode='generate'
+}
+async function queueCaption(stillId:string){
+ if(!stillId){error.value='Choose a still to describe.';return false}
+ try{
+  await $fetch('/api/studio-2/caption',{method:'POST',body:{folderId:form.folderId,stillId,sourcePath:stillId,captionStyle:captionStyle.value}})
+  await refresh()
+  return true
+ }catch(e:any){error.value=e.data?.statusMessage||e.message;return false}
+}
+function describeSelected(){
+ const id=selected.value?.kind==='still'?selected.value.id:(selected.value?.kind==='clip'? '' : (selected.value?.id || form.imageAId || form.identityStillId))
+ if(!id){error.value='Select a still to describe.';return}
+ captionStillId.value=id
+ submitting.value=true;error.value=''
+ void queueCaption(id).finally(()=>{submitting.value=false})
+}
+async function uploadCaption(event:Event){
+ const input=event.target as HTMLInputElement,file=input.files?.[0];if(!file)return
+ uploading.value=true;uploadMessage.value='';error.value=''
+ const body=new FormData();body.append('image',file);body.append('folderId',form.folderId);body.append('hideInput',String(xaigen))
+ try{const response=await $fetch<any>('/api/studio-2/stills',{method:'POST',body});library.value=response;captionStillId.value=response.still.id;measured[response.still.id]={width:response.still.width,height:response.still.height};uploadMessage.value=`Loaded ${response.still.name || file.name} for caption.`}
+ catch(e:any){error.value=e.data?.statusMessage||e.message}
+ finally{uploading.value=false;input.value=''}
+}
+
+function isHidden(asset:any){
+ if(!xaigen || !asset?.id)return false
+ if(isClassic.value){
+  const input=asset.role==='input'
+  return input?classicHideInput.value:classicHideOutput.value
+ }
+ return previewHidden(asset,revealed.value,xaigen)
+}
+function slotHidden(id:string){
+ if(!xaigen || !id)return false
+ if(isClassic.value)return classicHideInput.value
+ return !revealed.value.has(id)
+}
+function toggleReveal(id:string){
+ if(!id)return
+ const next=new Set(revealed.value)
+ if(next.has(id))next.delete(id);else next.add(id)
+ revealed.value=next
+ if(import.meta.client)localStorage.setItem(`studio2-revealed:${form.folderId}`,JSON.stringify([...next]))
+}
+function loadRevealed(folderId:string){
+ if(!import.meta.client||!xaigen){revealed.value=new Set();return}
+ try{revealed.value=new Set(JSON.parse(localStorage.getItem(`studio2-revealed:${folderId}`)||'[]'))}
+ catch{revealed.value=new Set()}
+}
+
 watch(()=>sourceAsset.value?.id,id=>{const asset=sourceAsset.value;if(!import.meta.client||!asset||measured[id])return;const media=asset.id===form.startClipId?document.createElement('video'):new Image();if(media.tagName==='VIDEO')(media as HTMLVideoElement).preload='metadata';const done=()=>{measured[asset.id]={width:(media as HTMLVideoElement).videoWidth||(media as HTMLImageElement).naturalWidth,height:(media as HTMLVideoElement).videoHeight||(media as HTMLImageElement).naturalHeight}};media.addEventListener(media.tagName==='VIDEO'?'loadedmetadata':'load',done,{once:true});media.src=media.tagName==='VIDEO'?`/api/library/clips/${asset.id}/video`:stillUrl(asset.id)})
+
 function chooseTask(task:string){
+ if(task==='music'){form.mode='music';return}
+ if(task==='caption'){
+  form.mode='caption'
+  captionStillId.value=selected.value?.kind==='still'?selected.value.id:form.imageAId||form.identityStillId||captionStillId.value
+  return
+ }
  if(task==='extend'){
   form.mode='extend'
   form.engine=xaigen && form.engine==='ltx'?'ltx':'minimax'
@@ -106,42 +486,92 @@ function chooseTask(task:string){
   }
   return
  }
-  const was=video.value;form.mode=task;form.startClipId='';if(video.value!==was){form.engine=video.value?'minimax':'flux';form.settings.steps=video.value?8:24;form.settings.cfg=video.value?1.5:4}
+ const was=video.value;form.mode=task;form.startClipId='';if(video.value!==was){form.engine=video.value?(xaigen && form.engine==='ltx'?'ltx':'minimax'):'flux';form.settings.steps=video.value?8:24;form.settings.cfg=video.value?1.5:4}
 }
+function onTaskSelect(event:Event){chooseTask((event.target as HTMLSelectElement).value)}
 function chooseClip(id:string){if(!id){form.startClipId='';return}const clip=clips.value.find((c:any)=>c.id===id);if(clip){try{Object.assign(form,bindClip(form,clip,xaigen,loraMemory))}catch(e:any){error.value=e?.message||'Could not copy clip settings.';form.startClipId=clip.id;form.mode='extend';form.engine=xaigen && form.engine==='ltx'?'ltx':'minimax'}}}
 function selectOutput(item:any){selected.value=item;if(form.mode==='extend' && item?.kind==='clip')chooseClip(item.id)}
-function sourceChanged(){Object.assign(form,changeSource(form))}
 function restore(item:any){Object.assign(form,restoreAll(form,item,xaigen,loraMemory));form.guides=form.guides||[];batchText.value=''}
 function focusJob(job:any){selectOutput(job.outputs?.at(-1)||null)}
 async function loadLoras(fresh=false){try{loraNames.value=await $fetch('/api/studio-2/loras'+(fresh?'?fresh=1':''))}catch{error.value='Could not refresh LoRAs. The last loaded list is still available.'}}
-function pickAsset(asset:any){const key=picker.value;if(key.startsWith('guide:'))form.guides[Number(key.split(':')[1])].stillId=asset?.id||'';else if(key==='startClipId')chooseClip(asset?.id||'');else form[key]=asset?.id||'';picker.value=''}
+function pickAsset(asset:any){const key=picker.value;if(key.startsWith('guide:'))form.guides[Number(key.split(':')[1])].stillId=asset?.id||'';else if(key==='startClipId')chooseClip(asset?.id||'');else if(key==='captionStillId')captionStillId.value=asset?.id||'';else form[key]=asset?.id||'';picker.value=''}
 async function clearFinished(){await $fetch('/api/studio-2/clear-finished',{method:'POST'});await refresh()}
-async function deleteOutput(){if(!selected.value)return;try{await $fetch('/api/library/'+(selected.value.kind==='clip'?'clips/':'stills/')+selected.value.id,{method:'DELETE'});await refresh()}catch(e:any){error.value=e.data?.statusMessage||e.message}}
-async function queue(){submitting.value=true;error.value='';try{if(form.settings.seedMode!=='locked')form.settings.seed=randomSeed();const queued=await $fetch<any>('/api/studio-2/jobs',{method:'POST',body:{...form,mode:form.mode==='edit' && parseBatch(batchText.value).length?'iterate':form.mode,hideInput:inputHidden.value,hideOutput:xaigen && hideOutput.value,promptSections:video.value?form.promptSections:{action:form.promptSections.action},settings:{...form.settings,...resolvedSize.value},batch:video.value || form.mode==='edit'?parseBatch(batchText.value):[]}});form.settings.seed=queued.seed;form.settings.width=queued.width;form.settings.height=queued.height;await refresh()}catch(e:any){error.value=e.data?.statusMessage||e.message}finally{submitting.value=false}}
+async function deleteOutput(){
+ if(!selected.value)return
+ try{
+  const kind=selected.value.kind==='clip'?'clips/':selected.value.kind==='track'?'tracks/':'stills/'
+  await $fetch('/api/library/'+kind+selected.value.id,{method:'DELETE'})
+  await refresh()
+ }catch(e:any){error.value=e.data?.statusMessage||e.message}
+}
+async function queue(){
+ submitting.value=true;error.value=''
+ try{
+  if(form.mode==='music'){
+   await $fetch('/api/generate/music',{method:'POST',body:{engine:'yue2',tags:music.tags,lyrics:music.lyrics,duration:music.duration,seed:music.seed||'random',folderId:form.folderId,name:music.name}})
+   await refresh()
+   return
+  }
+  if(form.mode==='caption'){
+   await queueCaption(captionStillId.value || form.imageAId || form.identityStillId)
+   return
+  }
+  if(form.settings.seedMode!=='locked')form.settings.seed=randomSeed()
+  const queued=await $fetch<any>('/api/studio-2/jobs',{method:'POST',body:{...form,mode:form.mode==='edit' && parseBatch(batchText.value).length?'iterate':form.mode,hideInput:xaigen,hideOutput:false,promptSections:video.value?form.promptSections:{action:form.promptSections.action},settings:{...form.settings,...resolvedSize.value},batch:video.value || form.mode==='edit'?parseBatch(batchText.value):[]}})
+  form.settings.seed=queued.seed;form.settings.width=queued.width;form.settings.height=queued.height
+  await refresh()
+ }catch(e:any){error.value=e.data?.statusMessage||e.message}
+ finally{submitting.value=false}
+}
 async function cancel(job:any){await $fetch('/api/studio-2/cancel',{method:'POST',body:{id:job.id}});await refresh()}
 async function upload(event:Event,target:'source'|'hero'){
  const input=event.target as HTMLInputElement,file=input.files?.[0];if(!file)return
  uploading.value=true;uploadMessage.value='';error.value=''
- const body=new FormData();body.append('image',file);body.append('folderId',form.folderId);body.append('hideInput',String(form.hideInput))
+ const body=new FormData();body.append('image',file);body.append('folderId',form.folderId);body.append('hideInput',String(xaigen))
  try{const response=await $fetch<any>('/api/studio-2/stills',{method:'POST',body});library.value=response;Object.assign(form,applyUploadedStill(form,response,target));measured[response.still.id]={width:response.still.width,height:response.still.height};uploadMessage.value=`Loaded ${response.still.name || file.name} as ${target==='hero'?'hero identity':'start still'}.`}
  catch(e:any){error.value=e.data?.statusMessage||e.message}
  finally{uploading.value=false;input.value=''}
 }
 
 async function refresh(){
- const [nextJobs,nextLibrary,nextUpscales]=await Promise.all([$fetch<any[]>('/api/studio-2/jobs'),$fetch<any>('/api/studio-2/library'),$fetch<any[]>('/api/studio-2/upscales')])
+ const [nextJobs,nextLibrary,nextUpscales,queue]=await Promise.all([
+  $fetch<any[]>('/api/studio-2/jobs'),
+  $fetch<any>('/api/studio-2/library'),
+  $fetch<any[]>('/api/studio-2/upscales'),
+  $fetch<any>('/api/studio-queue?full=1').catch(()=>({jobs:[]}))
+ ])
  jobs.value=nextJobs;library.value=nextLibrary;upscales.value=nextUpscales
+ studioQueue.value=Array.isArray(queue)?queue:(queue?.jobs||[])
+ const musicRows=studioQueue.value.filter((j:any)=>j.kind==='music' && j.liveJobId && !['complete','done'].includes(j.status))
+ const liveEntries=await Promise.all(musicRows.map(async(j:any)=>{
+  try{return [j.liveJobId,await $fetch<any>(`/api/generate/${j.liveJobId}`)] as const}
+  catch{return [j.liveJobId,null] as const}
+ }))
+ const nextLive:Record<string,any>={}
+ for(const [id,snap] of liveEntries){if(snap)nextLive[id]=snap}
+ liveMusic.value=nextLive
  const fresh=newestOutput(nextJobs,seenOutputs,form.folderId)
  for(const job of nextJobs)for(const output of job.outputs||[])seenOutputs.add(output.id)
  const existingFresh=family.value.find(a=>a.id===fresh?.id)
+ const newestTrack=(library.value.tracks||[]).filter((t:any)=>t.folderId===form.folderId).sort((a:any,b:any)=>(b.createdAt||0)-(a.createdAt||0))[0]
+ if(newestTrack && !seenOutputs.has(newestTrack.id)){seenOutputs.add(newestTrack.id);selected.value={...newestTrack,kind:'track'};return}
  selected.value=existingFresh || reconcileSelection(selected.value,family.value)
 }
 
-watch(()=>form.folderId,id=>{selected.value=null;form.imageAId='';form.startClipId='';if(import.meta.client)form.identityStillId=localStorage.getItem(`studio2-hero:${id}`)||''},{flush:'sync'})
+watch(()=>form.folderId,id=>{selected.value=null;form.imageAId='';form.startClipId='';loadRevealed(id);if(import.meta.client)form.identityStillId=localStorage.getItem(`studio2-hero:${id}`)||''},{flush:'sync'})
 watch(()=>form.identityStillId,id=>{if(import.meta.client&&form.folderId)localStorage.setItem(`studio2-hero:${form.folderId}`,id)})
-watch(hideOutput,value=>{if(import.meta.client)localStorage.setItem('studio2-hide-output',String(value))})
-watch(()=>form.hideInput,value=>{if(import.meta.client)localStorage.setItem('studio2-hide-input',String(value))})
 let timer:ReturnType<typeof setTimeout>,clock:ReturnType<typeof setInterval>,disposed=false
-onMounted(async()=>{let storage;try{storage=localStorage}catch{};loraMemory=createLoraMemory(xaigen,storage);loraMemory.sync(form);form.settings.seed=randomSeed();form.hideInput=xaigen && localStorage.getItem('studio2-hide-input')==='true';hideOutput.value=xaigen && localStorage.getItem('studio2-hide-output')==='true';clock=setInterval(()=>{now.value=Date.now()},1000);void loadLoras();library.value=await $fetch('/api/studio-2/library');form.folderId=library.value.folders[0]?.id||'';const id=String(route.query.still||route.query.clip||'');const asset=[...library.value.stills,...library.value.clips].find((a:any)=>a.id===id);if(asset){selected.value={...asset,kind:route.query.clip?'clip':'still'};if(route.query.use==='1')restore(selected.value)};const poll=async()=>{try{await refresh();const health=await $fetch<any>('/api/health');healthState.value=health}catch(e:any){healthState.value=null;error.value=e.data?.statusMessage||e.message}if(!disposed)timer=setTimeout(poll,3000)};void poll()})
+onMounted(async()=>{
+ let storage;try{storage=localStorage}catch{}
+ loraMemory=createLoraMemory(xaigen,storage);loraMemory.sync(form);form.settings.seed=randomSeed()
+ clock=setInterval(()=>{now.value=Date.now()},1000);void loadLoras(true)
+ library.value=await $fetch('/api/studio-2/library');form.folderId=library.value.folders[0]?.id||'';loadRevealed(form.folderId)
+ const id=String(route.query.still||route.query.clip||route.query.track||'')
+ const asset=[...library.value.stills,...library.value.clips,...(library.value.tracks||[])].find((a:any)=>a.id===id)
+ if(asset){selected.value={...asset,kind:route.query.clip?'clip':route.query.track?'track':'still'};if(route.query.use==='1'&&selected.value.kind!=='track')restore(selected.value);if(xaigen)toggleReveal(asset.id)}
+ if(route.query.task==='music')form.mode='music'
+ if(route.query.task==='caption')form.mode='caption'
+ const poll=async()=>{try{await refresh();const health=await $fetch<any>('/api/health');healthState.value=health}catch(e:any){healthState.value=null;error.value=e.data?.statusMessage||e.message}if(!disposed)timer=setTimeout(poll,3000)};void poll()
+})
 onBeforeUnmount(()=>{disposed=true;clearTimeout(timer);clearInterval(clock)})
 </script>

composables/useDisplayMode.ts

@@ -0,0 +1,82 @@
+export type DisplayMode = 'classic' | 'studio'
+
+const STORAGE_KEY = 'aigen.displayMode'
+
+function parseMode(raw: unknown): DisplayMode | null {
+  const value = String(raw || '').trim().toLowerCase()
+  return value === 'classic' || value === 'studio' ? value : null
+}
+
+function readStored(): DisplayMode {
+  if (!import.meta.client) return 'studio'
+  try {
+    return parseMode(localStorage.getItem(STORAGE_KEY)) || 'studio'
+  } catch {
+    return 'studio'
+  }
+}
+
+function writeStored(mode: DisplayMode) {
+  if (!import.meta.client) return
+  try {
+    localStorage.setItem(STORAGE_KEY, mode)
+  } catch { /* ignore quota / private mode */ }
+}
+
+/** Classic | Studio chrome skin. One Bench tree; layout only. */
+export function useDisplayMode() {
+  const route = useRoute()
+  const displayMode = useState<DisplayMode>('aigen.displayMode', () => 'studio')
+  const sessionOverride = useState<DisplayMode | null>('aigen.displayMode.session', () => null)
+  const hydrated = useState('aigen.displayMode.hydrated', () => false)
+
+  const resolved = computed<DisplayMode>(() => {
+    const fromQuery = parseMode(route.query.display)
+    if (fromQuery) return fromQuery
+    if (sessionOverride.value) return sessionOverride.value
+    return displayMode.value
+  })
+
+  function setDisplayMode(mode: DisplayMode, opts?: { persist?: boolean; session?: boolean }) {
+    const persist = opts?.persist !== false
+    if (opts?.session) sessionOverride.value = mode
+    else sessionOverride.value = null
+    displayMode.value = mode
+    if (persist) writeStored(mode)
+  }
+
+  function hydrate() {
+    if (hydrated.value) return
+    hydrated.value = true
+    const fromQuery = parseMode(route.query.display)
+    if (fromQuery) {
+      sessionOverride.value = fromQuery
+      return
+    }
+    displayMode.value = readStored()
+  }
+
+  if (import.meta.client) {
+    onMounted(hydrate)
+    watch(
+      () => route.query.display,
+      (raw) => {
+        const fromQuery = parseMode(raw)
+        if (fromQuery) sessionOverride.value = fromQuery
+        else if (sessionOverride.value && !parseMode(route.query.display)) {
+          // Drop session override when query cleared; keep stored preference.
+          sessionOverride.value = null
+          displayMode.value = readStored()
+        }
+      }
+    )
+  }
+
+  return {
+    displayMode: resolved,
+    storedDisplayMode: displayMode,
+    setDisplayMode,
+    isStudio: computed(() => resolved.value === 'studio'),
+    isClassic: computed(() => resolved.value === 'classic')
+  }
+}

docs/caption.md

@@ -0,0 +1,40 @@
+# Studio 2 image→text caption (Qwen2.5-VL NSFW Caption V4 GGUF)
+
+Caption jobs run on the Windows GPU host via **llama-server** (llama.cpp vision / mmproj). Not Comfy, not Klein, not JoyCaption.
+
+## Weights (~7GB — two files only)
+
+```powershell
+powershell -ExecutionPolicy Bypass -File scripts\setup-caption.ps1
+```
+
+Installs into the Shared models tree:
+
+`%LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Shared\models\caption\qwen25vl-7b-nsfw-v4\`
+
+| File | Role |
+| --- | --- |
+| `Qwen2.5-VL-7B-NSFW-Caption-V4.Q5_K_M.gguf` | Language model |
+| `Qwen2.5-VL-7B-NSFW-Caption-V4.mmproj-f16.gguf` | Vision projector |
+
+Do **not** download the full 16.6GB safetensors repo.
+
+Runtime needs ~8–10GB VRAM. Exclusive GPU: host stops Comfy (and waits if YuE2/upscale is busy), loads the GGUF for one shot, then kills llama-server (`keep_alive 0`).
+
+## Host agent
+
+Requires `llama-server` on PATH (`winget install ggml.llamacpp`) or `CAPTION_LLAMA_SERVER`.
+
+Restart the Comfy host agent after setup. Confirm:
+
+```text
+GET http://127.0.0.1:8199/caption/status  →  { configured, busy, backend: "llama.cpp" }
+POST /caption  { imagePath, style }       →  load → { text } → unload
+POST /caption/jobs + PUT …/input          →  async job used by Studio 2
+```
+
+Styles: `descriptive` | `klein_prompt` | `delta` | `tags`.
+
+## App
+
+`POST /api/studio-2/caption` with `{ folderId, stillId|sourcePath, captionStyle }` queues a Studio job. Bench **Describe** button and the **caption** task pill both enqueue. Result stores `resultText` on the studio2 job and optionally writes `{still}.txt` beside the library file.

docs/qwen21.md

@@ -0,0 +1,42 @@
+# Qwen Image 2.1 (engine `qwen21`) — host weights + Comfy graph
+
+## Host paths (this RTX 5080 box)
+
+| Role | Resolved path |
+|---|---|
+| ComfyUI root (Klein / host agent `ComfyUI (1)`) | `C:\Users\ianjm\AppData\Local\Comfy-Desktop\ComfyUI-Installs\ComfyUI (1)\ComfyUI` |
+| Models root (Desktop Shared via `shared_model_paths.yaml`) | `C:\Users\ianjm\AppData\Local\Comfy-Desktop\ComfyUI-Shared\models` |
+
+Weights must land under Shared so every Desktop instance sees them:
+
+- `diffusion_models\qwen-image-2.1-Q8_0.gguf` (~7.59 GiB) — from `abenzerps/Qwen-Image-2.1-Uncensored-GGUF`
+- `text_encoders\qwen3vl_8b_int8_convrot.safetensors` — from `Comfy-Org/Qwen-Image-2.1` (INT8; BF16 8B VL OOMs on 16 GB)
+- `vae\qwen_image_2.1_vae_bf16.safetensors` — from `Comfy-Org/Qwen-Image-2.1` (**not** the old Qwen-Image 1.0 VAE)
+
+Custom node: `custom_nodes\ComfyUI-GGUF` (`UnetLoaderGGUF`). Do not install a second GGUF pack.
+
+## Setup
+
+```powershell
+powershell -ExecutionPolicy Bypass -File scripts\setup-qwen21.ps1
+```
+
+DiT download uses the exact host invocation:
+
+```powershell
+hf download hf://abenzerps/Qwen-Image-2.1-Uncensored-GGUF/qwen-image-2.1-Q8_0.gguf
+```
+
+That lands in the Hugging Face hub cache; the setup script copies it into Shared `diffusion_models`. The abenzerps GGUF ships with `kv_count=0` (no `general.architecture`) **and** Q8_0-quantized 1D RMSNorm weights (logical 128 → packed 136), which breaks Comfy’s fused `rms_rope`. `scripts/tag-qwen21-gguf.py` rewrites the file with `general.architecture=qwen_image` and promotes small/1D tensors to F32 so city96 `UnetLoaderGGUF` can load it.
+
+Restart Comfy **only when idle** (`COMFY_CONTROL_URL/status` → `gpu.busy=false`). Confirm `object_info` lists `UnetLoaderGGUF` and the three filenames.
+
+## App
+
+- Engine key: `qwen21` · UI label: **Qwen 2.1**
+- Generate (T2I) only in this build. Edit / Compose / Iterate / Video disable with: “Qwen 2.1 is T2I in this build”.
+- Sampler defaults: euler / simple / cfg **1** / steps **25** · `ModelSamplingAuraFlow` shift **3.1**
+- Default canvas **1024×1024** (aspect 16:9 / 9:16 / 1:1 → long-edge square via `TextEncodeQwenImage21` resolution; multiples of 32). Drop to 768 if VRAM errors.
+- No hero / locks / Klein LoRA stack on this engine.
+
+Graph: `server/assets/studio2_qwen21_t2i.json`.

docs/yue2.md

@@ -1,6 +1,26 @@
 # YuE2 music engine
 
-`engine: "yue2"` is standalone [YuE2](https://github.com/multimodal-art-projection/YuE) (`m-a-p/YuE2-3B` + `m-a-p/YuE2-Vae`). It never submits a Comfy graph. `engine: "yue"` remains YuEGP in code. ACE / ACE 1.5 keep their Comfy workflows.
+`engine: "yue2"` is the only music engine in the app. Standalone [YuE2](https://github.com/multimodal-art-projection/YuE) (`m-a-p/YuE2-3B` + `m-a-p/YuE2-Vae`). It never submits a Comfy graph. ACE / YuEGP are removed from the picker and generate API; leave ACE Comfy checkpoints on disk.
+
+## Desktop agent at boot (one click)
+
+Wrong agent on 8199 → `/yue2/status` returns `{"ok":false,"error":"not found"}` while Comfy images still work.
+
+Fix once from this clone:
+
+```powershell
+.\scripts\install-comfy-host-agent-launchers.ps1
+```
+
+That rewrites Desktop, Start Menu, and **Windows Startup** shortcuts to this clone, and retargets the `AIGen Comfy Host Agent` scheduled task when it exists. Each launch **kills** any old `comfy-host-agent` / listener on 8199, then starts `scripts/comfy-host-agent.mjs` (imports `yue2-host.mjs`).
+
+After install: double-click Desktop **Start Comfy Host Agent**, or reboot. Confirm:
+
+```text
+GET http://127.0.0.1:8199/yue2/status  →  200 { configured, busy, backend: "yue2" }
+```
+
+Do not guess which `node.exe` to kill — the launcher replaces it.
 
 ## What this host already has (5080 desktop)
 
@@ -8,7 +28,7 @@ Ran on this machine:
 
 1. `scripts/setup-yue2.ps1` → checkout + venv + CUDA torch + weight download
 2. `scripts/cleanup-yue-v1.ps1` → named YuE v1 / YuEGP weight folders removed
-3. Host agent code mounts `POST/GET /yue2/jobs` (same pattern as `/yuegp/jobs`)
+3. Host agent mounts `POST/GET /yue2/jobs`
 
 Paths written:
 
@@ -23,7 +43,7 @@ Paths written:
 
 `aigen-ready.json` points `model` / `vae` at those local folders so Generate does not need a first-click HF pull. CUDA check: `torch 2.11.0+cu128` on `NVIDIA GeForce RTX 5080`.
 
-After deploying this branch, restart the **desktop host connector** once so it loads `yue2-host.mjs`. Do not restart Coolify from the agent; the user triggers the existing xAIGen-dev Coolify rebuild.
+After deploying this branch, run the installer above (or the Desktop shortcut) once so 8199 serves YuE2. Do not restart Coolify from the agent; live apps that track `main` rebuild on push.
 
 ## Env vars
 
@@ -35,7 +55,7 @@ After deploying this branch, restart the **desktop host connector** once so it l
 | `YUE2_MODEL` | Local model dir or HF id |
 | `YUE2_VAE` | Local VAE dir or HF id |
 
-App uses existing `COMFY_CONTROL_URL` + token. One GPU owner via the shared reservation (Comfy stopped while YuE2 runs). One song. Unload after decode. OOM is visible; no fallback to `yue`.
+App uses existing `COMFY_CONTROL_URL` + token. One GPU owner via the shared reservation (Comfy stopped while YuE2 runs). One song. Unload after decode. OOM is visible; no ACE/YuEGP fallback.
 
 ## Call path
 
@@ -60,19 +80,21 @@ Duration 30–150 s (default 60). One non-empty lyric section (same validation a
 - `C:\Users\ianjm\Development\YuEGP\inference\xcodec_mini_infer`
 - `C:\Users\ianjm\Development\YuEGP\aigen-jobs`
 
-HF cache YuE-s1 / YuE-s2 / xcodec entries were already absent (skipped). ACE checkpoints (`ace_step_v1_3.5b.safetensors`, `ace_step_1.5_turbo_aio.safetensors`) and other Comfy model trees were left alone. `scripts/yuegp-*` and `server/utils/yueGp.ts` remain in git.
+HF cache YuE-s1 / YuE-s2 / xcodec entries were already absent (skipped). ACE checkpoints (`ace_step_v1_3.5b.safetensors`, `ace_step_1.5_turbo_aio.safetensors`) and other Comfy model trees were left alone on disk.
 
 ## Re-run setup (only if rebuilding the machine)
 
 ```powershell
 .\scripts\setup-yue2.ps1
 .\scripts\cleanup-yue-v1.ps1
+.\scripts\install-comfy-host-agent-launchers.ps1
 ```
 
 Setup reuses existing `models\YuE2-3B` / `YuE2-Vae` when present, then force-reinstalls CUDA torch after the editable package install.
 
 ## Validation
 
-- JS: `node --test tests/yue2.test.mjs tests/yuegp.test.mjs`
-- Python: `py -3 -m unittest tests.test_yue2_worker tests.test_yuegp_worker`
-- No full song was generated in this agent session.
+- JS: `node --test tests/yue2.test.mjs`
+- Python: `py -3 -m unittest tests.test_yue2_worker`
+- After agent restart: `GET /yue2/status` returns 200.
+- Do not claim a song generated unless one ran after the agent restart.

nuxt.config.ts

@@ -4,7 +4,7 @@ const xaigenStudio = /^xaigen$/i.test(String(process.env.NUXT_PUBLIC_STUDIO || '
 export default defineNuxtConfig({
   compatibilityDate: '2025-08-01',
   modules: ['@nuxtjs/tailwindcss'],
-  css: ['~/assets/css/main.css'],
+  css: ['~/assets/css/main.css', '~/assets/css/display.css'],
   app: {
     head: {
       title: 'AIGen — MiniMax H3',

pages/music.vue

@@ -1,7 +1,7 @@
 <template>
   <div class="min-h-dvh min-w-0">
     <AppHeader
-      :subtitle="advanced ? headerSubtitle : 'Create music from text'"
+      :subtitle="advanced ? 'YuE2 · lyrics-to-song' : 'Create music from text'"
       current="music"
       :queue-count="queueCount"
     />
@@ -10,9 +10,8 @@
       <StudioKindCards
         :advanced="advanced"
         current="music"
-        :music-engine="engineFamily"
+        music-engine="yue2"
         @select="goStudio"
-        @music-engine="selectEngineFamily"
       />
       <div class="flex items-center justify-between gap-3"><span class="rounded-full border border-amber-300/60 px-4 py-2 text-sm text-amber-100">From text</span><label class="flex items-center gap-2 text-sm text-zinc-300"><input v-model="advanced" type="checkbox" class="accent-amber-400">Advanced mode</label></div>
       <div class="grid min-w-0 gap-6 lg:grid-cols-[minmax(0,1fr)_minmax(0,1fr)] lg:items-stretch">
@@ -20,43 +19,15 @@
         <div>
           <h2 class="font-display text-2xl font-bold">Music Generator</h2>
           <p class="mt-1 text-sm text-zinc-400">
-            Tags describe style and voice. Lyrics use structure chips. Instrumental skips the vocal line.
+            Tags describe style and voice. Lyrics use structure chips. Multiple sections are OK.
           </p>
         </div>
 
         <GenerationSettings :advanced="advanced">
-          <label class="block space-y-2 text-sm"><span>Model</span><select :value="engineFamily" class="w-full rounded-xl border border-white/15 bg-zinc-900 p-3" @change="selectEngineFamily(($event.target as HTMLSelectElement).value as 'ace' | 'yue' | 'yue2')"><option value="ace">ACE-Step</option><option value="yue">YuE</option><option value="yue2">YuE2</option></select></label>
-          <label v-if="engineFamily === 'yue'" class="block space-y-2 text-sm">
-            <span>YuEGP profile</span>
-            <select v-model.number="yueProfile" class="w-full rounded-xl border border-white/15 bg-zinc-900 p-3">
-              <option :value="1">1 · Full quality · 16GB (default)</option>
-              <option :value="3">3 · Lower memory · Manual fallback</option>
-            </select>
-            <span class="block text-xs text-zinc-400">One lyric section. Compile off. Profile 3 is never selected automatically.</span>
-          </label>
-          <p v-else-if="engineFamily === 'yue2'" class="text-xs text-zinc-400">YuE2-3B · one lyric section · plan → synthesize → decode · no score editor</p>
+          <p class="text-xs text-zinc-400">YuE2-3B · lyrics with section tags · plan → synthesize → decode · no score editor</p>
           <div v-if="!advanced" class="space-y-4">
-            <label v-if="engineFamily === 'ace'" class="block text-sm">Steps<input v-model.number="steps" type="number" :min="stepsMin" :max="stepsMax" class="mt-2 w-full rounded-xl border border-white/15 bg-zinc-900 p-3"></label>
             <label class="block text-sm">Seed<input v-model="seed" placeholder="random" class="mt-2 w-full rounded-xl border border-white/15 bg-zinc-900 p-3"></label>
           </div>
-        <div v-if="engineFamily === 'ace'" class="flex items-center justify-between gap-3 rounded-2xl border border-white/10 bg-zinc-950/40 px-4 py-3 text-sm">
-          <span class="min-w-0">
-            <span class="flex items-center gap-2">
-              <span class="block font-medium text-zinc-200">ACE-Step 1.5</span>
-              <InfoTip
-                aria-label="What ACE-Step 1.5 does"
-                :text="ace15Info"
-              />
-            </span>
-            <span class="mt-0.5 block text-xs text-zinc-500">Turbo AIO checkpoint. Different graph than v1, usually 8 steps.</span>
-          </span>
-          <label class="relative inline-flex h-6 w-11 shrink-0 cursor-pointer items-center" title="Toggle ACE-Step 1.5">
-            <input v-model="ace15" type="checkbox" class="peer sr-only" role="switch">
-            <span class="absolute inset-0 rounded-full bg-zinc-700 transition peer-checked:bg-amber-400" />
-            <span class="absolute left-0.5 top-0.5 h-5 w-5 rounded-full bg-white transition peer-checked:translate-x-5" />
-          </label>
-        </div>
-
         </GenerationSettings>
         <p v-if="error" class="rounded-2xl border border-red-400/30 bg-red-400/10 px-4 py-3 text-sm text-red-200">{{ error }}</p>
 
@@ -73,18 +44,13 @@
         <div>
           <div class="mb-1 flex flex-wrap items-center justify-between gap-2">
             <span class="text-sm font-medium text-zinc-300">Lyrics</span>
-            <label class="flex cursor-pointer items-center gap-2 text-xs text-zinc-400">
-              <input v-model="instrumental" :disabled="engineFamily === 'yue' || engineFamily === 'yue2'" type="checkbox" class="accent-amber-400">
-              Instrumental
-            </label>
           </div>
           <div class="mb-2 flex flex-wrap gap-1.5">
             <button
               v-for="chip in chips"
               :key="chip"
               type="button"
-              class="rounded-full border border-white/10 px-2.5 py-1 text-[11px] text-zinc-300 hover:border-amber-300/50 hover:text-white disabled:opacity-40"
-              :disabled="instrumental"
+              class="rounded-full border border-white/10 px-2.5 py-1 text-[11px] text-zinc-300 hover:border-amber-300/50 hover:text-white"
               @click="insertChip(chip)"
             >{{ chip }}</button>
           </div>
@@ -92,8 +58,7 @@
             ref="lyricsEl"
             v-model="lyrics"
             rows="12"
-            class="max-h-[min(28rem,50vh)] w-full overflow-y-auto rounded-2xl border border-white/10 bg-zinc-950/70 px-4 py-3 text-sm outline-none ring-amber-300/40 placeholder:text-zinc-600 [overflow-wrap:anywhere] focus:ring-2 disabled:opacity-50"
-            :disabled="instrumental"
+            class="max-h-[min(28rem,50vh)] w-full overflow-y-auto rounded-2xl border border-white/10 bg-zinc-950/70 px-4 py-3 text-sm outline-none ring-amber-300/40 placeholder:text-zinc-600 [overflow-wrap:anywhere] focus:ring-2"
             :placeholder="lyricsPlaceholder"
           />
         </div>
@@ -109,19 +74,7 @@
               :max="durationMax"
               step="5"
             >
-            <span class="mt-1 block text-[11px] text-zinc-500">{{ engineFamily === 'yue' || engineFamily === 'yue2' ? 'Approximate length · one lyric section' : `${durationMin}–${durationMax} seconds` }}</span>
-          </label>
-          <label v-if="advanced && engineFamily === 'ace'" class="block text-sm">
-            <span class="mb-1 block font-medium text-zinc-300">Steps · {{ steps }}</span>
-            <input
-              v-model.number="steps"
-              type="range"
-              class="aigen-slider w-full"
-              :min="stepsMin"
-              :max="stepsMax"
-              step="1"
-            >
-            <span class="mt-1 block text-[11px] text-zinc-500">{{ ace15 ? '8 is the usual 1.5 turbo range' : '30–50 is the usual range' }}</span>
+            <span class="mt-1 block text-[11px] text-zinc-500">Approximate length · multiple sections OK</span>
           </label>
         </div>
 
@@ -209,7 +162,7 @@
         </div>
         <p v-if="busy && !queued && jobId" class="text-xs text-zinc-400">
           {{ musicElapsed }} elapsed · {{ musicActivityText }}
-          <span v-if="(activeMusicEngine === 'yue' || activeMusicEngine === 'yue2') && !indeterminateMusic" class="mt-1 block">Progress is for the current stage.</span>
+          <span v-if="!indeterminateMusic" class="mt-1 block">Progress is for the current stage.</span>
           <span v-if="indeterminateMusic" class="mt-1 block">Waiting for a progress report.</span>
         </p>
         <AudioPlayer
@@ -236,18 +189,10 @@ const { advanced } = useStudioLayout()
 import {
   DEFAULT_MUSIC_LYRICS,
   DEFAULT_MUSIC_TAGS,
-  MUSIC_CFG_DEFAULT,
-  MUSIC_CFG_DEFAULT_15,
   MUSIC_DURATION_DEFAULT,
   MUSIC_DURATION_MAX,
   MUSIC_DURATION_MIN,
-  MUSIC_STEPS_DEFAULT,
-  MUSIC_STEPS_DEFAULT_15,
-  MUSIC_STEPS_MAX,
-  MUSIC_STEPS_MIN,
   MUSIC_STRUCTURE_CHIPS,
-  parseMusicEngine,
-  type MusicEngine,
   yueLyricsProblem,
 } from '~/utils/music'
 import { ensureGenerationNotifyPermission, notifyGenerationComplete } from '~/utils/generationNotify'
@@ -259,25 +204,16 @@ const tagsPlaceholder = DEFAULT_MUSIC_TAGS
 const lyricsPlaceholder = DEFAULT_MUSIC_LYRICS
 const chips = MUSIC_STRUCTURE_CHIPS
 const durationMin = MUSIC_DURATION_MIN
-const durationMax = computed(() => (engineFamily.value === 'yue' || engineFamily.value === 'yue2') ? 150 : MUSIC_DURATION_MAX)
-const stepsMin = MUSIC_STEPS_MIN
-const stepsMax = MUSIC_STEPS_MAX
-
-const ace15Info = 'Turns on ACE-Step 1.5 Turbo, the all-in-one checkpoint already on Beast. 1.5 uses a different Comfy graph than v1: a 1.5 text encoder (tags, BPM, key, duration) and EmptyAceStep1.5LatentAudio, usually with 8 sampler steps and CFG 1. Songs often come back faster and cleaner. Lyrics-strength from v1 does not apply. Leave this off to stay on ACE-Step v1 3.5B, which is the graph that already generated on this GPU.'
+const durationMax = MUSIC_DURATION_MAX
 
 const tags = ref(DEFAULT_MUSIC_TAGS)
 const lyrics = ref(DEFAULT_MUSIC_LYRICS)
-const instrumental = ref(false)
 const duration = ref(MUSIC_DURATION_DEFAULT)
-const steps = ref(MUSIC_STEPS_DEFAULT)
 const seed = ref('')
 const name = ref('')
 const folderId = ref('')
 const folders = ref<LibraryFolder[]>([])
 const lyricsEl = ref<HTMLTextAreaElement | null>(null)
-const ace15 = ref(false)
-const engineFamily = ref<'ace' | 'yue' | 'yue2'>('ace')
-const yueProfile = ref<1 | 3>(1)
 const busy = ref(false)
 const forceClearing = ref(false)
 const error = ref('')
@@ -285,10 +221,9 @@ const status = ref('')
 const progress = ref(0)
 const musicClock = ref(Date.now())
 const musicStartedAt = ref(0)
-const activeMusicEngine = ref('')
 const musicActivity = ref<{ checkedAt: number; running: boolean } | null>(null)
 let musicClockTimer: ReturnType<typeof setInterval> | null = null
-const indeterminateMusic = computed(() => !queued.value && (activeMusicEngine.value === 'yue' || activeMusicEngine.value === 'yue2') && progress.value < 98 && !musicActivity.value?.running)
+const indeterminateMusic = computed(() => !queued.value && progress.value < 98 && !musicActivity.value?.running)
 const recentMusicCheck = computed(() => Boolean(musicActivity.value?.running && musicClock.value - musicActivity.value.checkedAt < 15000))
 const musicElapsed = computed(() => {
   const seconds = musicStartedAt.value ? Math.max(0, Math.floor((musicClock.value - musicStartedAt.value) / 1000)) : 0
@@ -318,63 +253,19 @@ let queuePoll: ReturnType<typeof setInterval> | null = null
 let recoverPoll: ReturnType<typeof setInterval> | null = null
 
 const blocked = computed(() => Boolean(blockReason.value))
-const headerSubtitle = computed(() => {
-  if (engineFamily.value === 'yue2') return 'YuE2 · lyrics-to-song'
-  if (engineFamily.value === 'yue') return 'YuE · lyrics-to-song'
-  return ace15.value ? 'ACE-Step 1.5 · vocals' : 'ACE-Step · vocals'
-})
-const selectedEngine = computed<MusicEngine>(() => {
-  if (engineFamily.value === 'yue2') return 'yue2'
-  if (engineFamily.value === 'yue') return 'yue'
-  return ace15.value ? 'ace-step-1.5' : 'ace-step'
-})
 const blockReason = computed(() => {
   if (!folderId.value) return 'Choose a library folder before generating.'
-  if ((engineFamily.value === 'yue' || engineFamily.value === 'yue2') && instrumental.value) return 'YuE requires lyrics. Use ACE for instrumental music.'
-  if ((engineFamily.value === 'yue' || engineFamily.value === 'yue2') && duration.value > 150) return 'YuE supports up to 150 seconds per section.'
+  if (duration.value > 150) return 'YuE2 supports up to 150 seconds.'
   if (!tags.value.trim()) return 'Add genre and style tags.'
-  if (engineFamily.value === 'yue' || engineFamily.value === 'yue2') { const problem = yueLyricsProblem(lyrics.value); if (problem) return problem }
-  if (!instrumental.value && !lyrics.value.trim()) return 'Write lyrics, or turn on Instrumental.'
+  const problem = yueLyricsProblem(lyrics.value)
+  if (problem) return problem
   return ''
 })
 
-function selectEngineFamily(family: 'ace' | 'yue' | 'yue2') {
-  engineFamily.value = family
-  if (family === 'yue' || family === 'yue2') { instrumental.value = false; duration.value = Math.min(150, duration.value) }
-  if ((family === 'yue' || family === 'yue2') && lyrics.value === DEFAULT_MUSIC_LYRICS) lyrics.value = '[Verse 1]\n'
-  if (family === 'ace' && ace15.value && steps.value === MUSIC_STEPS_DEFAULT) {
-    steps.value = MUSIC_STEPS_DEFAULT_15
-  }
-  if (family === 'ace' && !ace15.value && steps.value === MUSIC_STEPS_DEFAULT_15) {
-    steps.value = MUSIC_STEPS_DEFAULT
-  }
-}
-
 function goStudio(kind: 'video' | 'image') {
   void navigateTo({ path: '/', query: { studio: kind } })
 }
 
-function applyEngine(engine: MusicEngine) {
-  if (engine === 'yue2') {
-    engineFamily.value = 'yue2'
-    ace15.value = false
-    return
-  }
-  if (engine === 'yue') {
-    engineFamily.value = 'yue'
-    ace15.value = false
-    return
-  }
-  engineFamily.value = 'ace'
-  ace15.value = engine === 'ace-step-1.5'
-}
-
-watch(ace15, (on) => {
-  if (engineFamily.value !== 'ace') return
-  if (on && (steps.value === MUSIC_STEPS_DEFAULT || steps.value >= 30)) steps.value = MUSIC_STEPS_DEFAULT_15
-  if (!on && steps.value === MUSIC_STEPS_DEFAULT_15) steps.value = MUSIC_STEPS_DEFAULT
-})
-
 function insertChip(chip: string) {
   const el = lyricsEl.value
   const token = `${chip}\n`
@@ -460,11 +351,7 @@ async function attachLiveMusic(opts: {
   if (typeof opts.progress === 'number') progress.value = opts.progress
   if (opts.message) status.value = opts.message
   else if (!status.value || /queueing|waiting for a generate|waiting in the job queue|waiting for gpu/i.test(status.value)) {
-    status.value = selectedEngine.value === 'yue'
-      ? 'YuEGP is generating your song'
-      : selectedEngine.value === 'yue2'
-        ? 'YuE2 is generating your song'
-        : 'Generating…'
+    status.value = 'YuE2 is generating your song'
   }
   if (jobId.value !== liveId) {
     musicActivity.value = null
@@ -507,11 +394,7 @@ async function refreshStudioQueue() {
     busy.value = true
     status.value = musicRow.status === 'held'
       ? 'Paused in the job queue'
-      : (selectedEngine.value === 'yue'
-        ? 'Waiting for the GPU to run YuEGP'
-        : selectedEngine.value === 'yue2'
-          ? 'Waiting for the GPU to run YuE2'
-          : 'Waiting in the job queue…')
+      : 'Waiting for the GPU to run YuE2'
   }
 
   if (!studioJobId.value) return
@@ -521,11 +404,7 @@ async function refreshStudioQueue() {
       queued.value = false
     }
     if (jobId.value && /waiting in the job queue/i.test(status.value)) {
-      status.value = selectedEngine.value === 'yue'
-        ? 'YuEGP is generating your song'
-        : selectedEngine.value === 'yue2'
-          ? 'YuE2 is generating your song'
-          : 'Generating…'
+      status.value = 'YuE2 is generating your song'
     }
     return
   }
@@ -534,11 +413,7 @@ async function refreshStudioQueue() {
     busy.value = true
     status.value = row.status === 'held'
       ? 'Paused in the job queue'
-      : (selectedEngine.value === 'yue'
-        ? 'Waiting for the GPU to run YuEGP'
-        : selectedEngine.value === 'yue2'
-          ? 'Waiting for the GPU to run YuE2'
-          : 'Waiting in the job queue…')
+      : 'Waiting for the GPU to run YuE2'
     return
   }
   if (row.status === 'running' && row.liveJobId) {
@@ -573,7 +448,6 @@ async function resumeActiveMusic() {
 
 function applyEvent(payload: Record<string, any>) {
   if (typeof payload.elapsedMs === 'number') musicStartedAt.value = Date.now() - payload.elapsedMs
-  if (payload.engine) activeMusicEngine.value = payload.engine
   if (payload.musicActivity) musicActivity.value = payload.musicActivity
   if (payload.message) status.value = payload.message
   if (typeof payload.progress === 'number') progress.value = payload.progress
@@ -681,13 +555,7 @@ async function generate() {
   trackId.value = ''
   jobId.value = ''
   studioJobId.value = ''
-  status.value = engineFamily.value === 'yue'
-    ? 'Queueing YuE…'
-    : engineFamily.value === 'yue2'
-      ? 'Queueing YuE2…'
-      : ace15.value
-        ? 'Queueing ACE-Step 1.5…'
-        : 'Queueing ACE-Step…'
+  status.value = 'Queueing YuE2…'
   try {
     const started = await $fetch<{ jobId: string; studioJobId?: string; queued?: boolean; name?: string }>(
       '/api/generate/music',
@@ -697,14 +565,10 @@ async function generate() {
           tags: tags.value,
           lyrics: lyrics.value,
           duration: duration.value,
-          steps: steps.value,
           seed: seed.value || 'random',
-          cfg: selectedEngine.value === 'ace-step-1.5' ? MUSIC_CFG_DEFAULT_15 : MUSIC_CFG_DEFAULT,
-          instrumental: instrumental.value,
           folderId: folderId.value,
           name: name.value.trim(),
-          engine: selectedEngine.value,
-          ...(engineFamily.value === 'yue' ? { yueProfile: yueProfile.value } : {})
+          engine: 'yue2'
         }
       }
     )
@@ -730,10 +594,6 @@ async function generate() {
   }
 }
 
-watch(instrumental, (on) => {
-  if (on && !lyrics.value.trim()) lyrics.value = DEFAULT_MUSIC_LYRICS
-})
-
 function applyIncomingMusic() {
   try {
     const genRaw = sessionStorage.getItem(LOAD_GENERATION_PAYLOAD_KEY)
@@ -744,14 +604,10 @@ function applyIncomingMusic() {
         const payload = loaded.payload
         if (typeof payload.prompt === 'string') tags.value = payload.prompt
         if (typeof payload.lyrics === 'string') lyrics.value = payload.lyrics
-        instrumental.value = payload.instrumental === true
-        if (typeof payload.duration === 'number') duration.value = payload.duration
-        if (typeof payload.steps === 'number') steps.value = payload.steps
+        if (typeof payload.duration === 'number') duration.value = Math.min(MUSIC_DURATION_MAX, payload.duration)
         if (payload.seed != null) seed.value = String(payload.seed)
         if (typeof payload.name === 'string') name.value = payload.name
         if (payload.folderId && folders.value.some(folder => folder.id === payload.folderId)) folderId.value = payload.folderId
-        if (payload.yueProfile === 1 || payload.yueProfile === 3) yueProfile.value = payload.yueProfile
-        if (payload.musicEngine || payload.engine) applyEngine(parseMusicEngine(payload.musicEngine || payload.engine))
         status.value = 'Loaded from generation log'
         return
       }
@@ -766,24 +622,16 @@ function applyIncomingMusic() {
       name?: string
       tags?: string
       lyrics?: string
-      instrumental?: boolean
       duration?: number
-      steps?: number
       seed?: string
       folderId?: string
-      engine?: string
-      yueProfile?: 1 | 3
     }
     if (typeof loaded.tags === 'string') tags.value = loaded.tags
     if (typeof loaded.lyrics === 'string') lyrics.value = loaded.lyrics
-    instrumental.value = loaded.instrumental === true
-    if (typeof loaded.duration === 'number') duration.value = loaded.duration
-    if (typeof loaded.steps === 'number') steps.value = loaded.steps
+    if (typeof loaded.duration === 'number') duration.value = Math.min(MUSIC_DURATION_MAX, loaded.duration)
     if (typeof loaded.seed === 'string') seed.value = loaded.seed
     if (typeof loaded.name === 'string') name.value = loaded.name
     if (loaded.folderId && folders.value.some(folder => folder.id === loaded.folderId)) folderId.value = loaded.folderId
-    if (loaded.engine) applyEngine(parseMusicEngine(loaded.engine))
-    yueProfile.value = loaded.yueProfile === 3 ? 3 : 1
     status.value = 'Loaded from the queue'
   } catch { /* ignore */ }
 }
@@ -799,12 +647,8 @@ function applyPreset() {
   const settings = preset.settings as MusicPresetSettings
   if (typeof settings.tags === 'string') tags.value = settings.tags
   if (typeof settings.lyrics === 'string') lyrics.value = settings.lyrics
-  instrumental.value = settings.instrumental === true
-  if (typeof settings.duration === 'number') duration.value = settings.duration
-  if (typeof settings.steps === 'number') steps.value = settings.steps
+  if (typeof settings.duration === 'number') duration.value = Math.min(MUSIC_DURATION_MAX, settings.duration)
   if (typeof settings.seed === 'number') seed.value = String(settings.seed)
-  if (settings.engine) applyEngine(settings.engine)
-  yueProfile.value = settings.yueProfile === 3 ? 3 : 1
   name.value = preset.name
 }
 
@@ -812,14 +656,12 @@ onMounted(async () => {
   musicClockTimer = setInterval(() => { musicClock.value = Date.now() }, 1000)
   await loadLibrary()
   await loadMusicPresets()
-  const route = useRoute()
-  const engineQ = String(route.query.engine || '').toLowerCase()
-  if (engineQ === 'yue' || engineQ === 'yue2' || engineQ === 'ace') selectEngineFamily(engineQ as 'ace' | 'yue' | 'yue2')
   applyIncomingMusic()
   await resumeActiveMusic()
   await refreshStudioQueue()
   stopQueuePoll()
   queuePoll = setInterval(() => { void refreshStudioQueue() }, 8000)
+  const route = useRoute()
   const query = route.query
   if (typeof query.track === 'string' && query.track) {
     audioUrl.value = `/api/library/tracks/${query.track}/audio`
@@ -833,7 +675,7 @@ onMounted(async () => {
       if (recovered?.trackId && /recovered/i.test(String(recovered.message || ''))) {
         applyEvent({ ...recovered, type: 'complete' })
       }
-    } catch { /* nothing waiting in ComfyUI */ }
+    } catch { /* nothing waiting */ }
   }
 })
 

pages/queue.vue

@@ -663,7 +663,7 @@ function progressLine(queue: Queue) {
 
 function jobLine(job: StudioJobRow) {
   if (job.kind === 'music') {
-    if (job.status === 'waiting') return job.cutIn ? 'Next after the current shot · ACE-Step' : 'Waiting · ACE-Step'
+    if (job.status === 'waiting') return job.cutIn ? 'Next after the current shot · YuE2' : 'Waiting · YuE2'
     if (job.status === 'held') return 'Paused'
     if (job.status === 'error') return job.lastError || 'Failed'
     return 'Generating track'
@@ -1237,7 +1237,7 @@ async function clearAll() {
 async function forceClearAll() {
   if (forceClearing.value) return
   if (!confirm(
-    'Force clear cancels every running and waiting job (including stuck YuE/music), wipes shot batches, and kills Comfy if it is wedged. Saved library clips stay. Continue?'
+    'Force clear cancels every running and waiting job (including stuck YuE2/music), wipes shot batches, and kills Comfy if it is wedged. Saved library clips stay. Continue?'
   )) return
   forceClearing.value = true
   error.value = ''

pages/studio-2.vue

@@ -1,7 +1,6 @@
 <template>
-  <div v-if="config.public.studio === 'xaigen'" class="border-b border-amber-400/20 bg-zinc-950 px-6 py-3 text-sm text-amber-100">xAIGen 2.0 development preview · Image and MiniMax tools are available here. <NuxtLink to="/" class="underline">Open the existing studio for LTX and music</NuxtLink>.</div>
-  <Studio2Bench />
+  <div />
 </template>
 <script setup lang="ts">
-const config = useRuntimeConfig()
+await navigateTo('/', { replace: true })
 </script>

scripts/caption-host.mjs

@@ -0,0 +1,331 @@
+import { spawn } from 'node:child_process'
+import { createServer } from 'node:net'
+import { existsSync, mkdirSync, readFileSync, writeFileSync, renameSync, appendFileSync, readdirSync, copyFileSync } from 'node:fs'
+import { join, resolve, extname } from 'node:path'
+import { captionStylePrompt, CAPTION_STYLES } from '../shared/studio2/caption.mjs'
+
+const MODEL_NAME = 'Qwen2.5-VL-7B-NSFW-Caption-V4.Q5_K_M.gguf'
+const MMPROJ_NAME = 'Qwen2.5-VL-7B-NSFW-Caption-V4.mmproj-f16.gguf'
+
+function defaultModelsDir() {
+  const shared = process.env.COMFY_MODELS_ROOT
+    || join(process.env.LOCALAPPDATA || '', 'Comfy-Desktop', 'ComfyUI-Shared', 'models')
+  return resolve(process.env.CAPTION_MODELS || join(shared, 'caption', 'qwen25vl-7b-nsfw-v4'))
+}
+
+function defaultLlamaServer() {
+  if (process.env.CAPTION_LLAMA_SERVER) return resolve(process.env.CAPTION_LLAMA_SERVER)
+  return 'llama-server'
+}
+
+function freePort() {
+  return new Promise((resolvePort, reject) => {
+    const server = createServer()
+    server.listen(0, '127.0.0.1', () => {
+      const address = server.address()
+      const port = typeof address === 'object' && address ? address.port : 0
+      server.close(error => error ? reject(error) : resolvePort(port))
+    })
+    server.on('error', reject)
+  })
+}
+
+async function waitForServer(port, signal, timeoutMs = 120_000, fetchImpl = fetch) {
+  const started = Date.now()
+  while (Date.now() - started < timeoutMs) {
+    if (signal?.aborted) throw new Error('Caption cancelled while loading the model.')
+    try {
+      const response = await fetchImpl(`http://127.0.0.1:${port}/health`, { signal: AbortSignal.timeout(2000) })
+      if (response.ok) return
+    } catch { /* still booting */ }
+    await new Promise(r => setTimeout(r, 500))
+  }
+  throw new Error('llama-server did not become ready in time.')
+}
+
+function mimeFor(path) {
+  const ext = extname(path).toLowerCase()
+  if (ext === '.jpg' || ext === '.jpeg') return 'image/jpeg'
+  if (ext === '.webp') return 'image/webp'
+  if (ext === '.gif') return 'image/gif'
+  return 'image/png'
+}
+
+function killTree(child) {
+  return new Promise(resolveKill => {
+    if (!child?.pid) return resolveKill()
+    let finished = false
+    const done = () => {
+      if (finished) return
+      finished = true
+      resolveKill()
+    }
+    child.once('close', done)
+    try {
+      if (typeof child.kill === 'function') child.kill()
+    } catch { /* ignore */ }
+    if (process.platform === 'win32' && child.pid > 0) {
+      try {
+        const killer = spawn('taskkill', ['/PID', String(child.pid), '/T', '/F'], { windowsHide: true, stdio: 'ignore' })
+        killer.once('close', done)
+        killer.once('error', done)
+      } catch {
+        done()
+      }
+    }
+    setTimeout(done, 3000)
+  })
+}
+
+export function validateCaptionHostRequest(body) {
+  if (!/^[a-zA-Z0-9-]{12,80}$/.test(body?.id || '')) throw new Error('Invalid caption job ID.')
+  const style = String(body.style || body.captionStyle || 'descriptive')
+  if (!CAPTION_STYLES.includes(style)) throw new Error('Unknown caption style.')
+  return { id: body.id, style, imagePath: body.imagePath ? String(body.imagePath) : '' }
+}
+
+/** One llama-server process per caption. Process exit is the VRAM unload boundary (keep_alive 0). */
+export function createCaptionHost({
+  prepare,
+  leaseValid,
+  spawnProcess = spawn,
+  modelsDir,
+  llamaServer,
+  dataDir,
+  fetchImpl = fetch
+} = {}) {
+  const models = resolve(modelsDir || defaultModelsDir())
+  const executable = llamaServer || defaultLlamaServer()
+  const data = resolve(dataDir || process.env.CAPTION_JOBS_DIR || join(models, 'aigen-jobs'))
+  let active = null
+  let holdUntil = 0
+  if (existsSync(data)) {
+    for (const id of readdirSync(data)) {
+      try {
+        const path = join(data, id, 'status.json')
+        if (!existsSync(path)) continue
+        const state = JSON.parse(readFileSync(path, 'utf8'))
+        if (['running', 'starting', 'cancelling'].includes(state.status)) {
+          holdUntil = Date.now() + 10000
+          state.status = 'error'
+          state.error = 'Caption host restarted. GPU model unloaded.'
+          writeFileSync(path + '.tmp', JSON.stringify(state))
+          renameSync(path + '.tmp', path)
+        }
+      } catch { /* ignore bad records */ }
+    }
+  }
+  const dir = id => {
+    if (!/^[a-zA-Z0-9-]{12,80}$/.test(id || '')) throw new Error('Invalid caption job ID.')
+    return join(data, id)
+  }
+  const persist = job => {
+    const target = join(dir(job.id), 'status.json')
+    writeFileSync(target + '.tmp', JSON.stringify(job))
+    renameSync(target + '.tmp', target)
+  }
+  const read = id => {
+    if (active?.job.id === id) return { ...active.job, text: active.job.text }
+    const path = join(dir(id), 'status.json')
+    if (!existsSync(path)) return null
+    return JSON.parse(readFileSync(path, 'utf8'))
+  }
+  const modelPath = () => join(models, MODEL_NAME)
+  const mmprojPath = () => join(models, MMPROJ_NAME)
+  const configured = () => {
+    try {
+      if (!existsSync(modelPath()) || !existsSync(mmprojPath())) return false
+      if (executable.includes('\\') || executable.includes('/')) return existsSync(executable)
+      return true
+    } catch { return false }
+  }
+
+  return {
+    busy: () => Boolean(active) || Date.now() < holdUntil,
+    configured,
+    modelsDir: () => models,
+    read,
+    async upload(id, stream) {
+      if (read(id)?.status === 'complete') throw new Error('This caption job has already finished.')
+      const jobDir = dir(id)
+      mkdirSync(jobDir, { recursive: true })
+      const dest = join(jobDir, 'input.upload')
+      const { createWriteStream } = await import('node:fs')
+      const { pipeline } = await import('node:stream/promises')
+      const { Transform } = await import('node:stream')
+      let bytes = 0
+      await pipeline(stream, new Transform({
+        transform(chunk, _, callback) {
+          bytes += chunk.length
+          callback(bytes > 40 * 1024 * 1024 ? new Error('Image exceeds the 40 MB upload limit.') : null, chunk)
+        }
+      }), createWriteStream(dest, { flags: 'w' }))
+      if (bytes < 32) throw new Error('Image file is empty.')
+      renameSync(dest, join(jobDir, 'input.png'))
+    },
+    async start(body, lease) {
+      const request = validateCaptionHostRequest(body)
+      const previous = read(request.id)
+      if (previous) return previous
+      if (active || Date.now() < holdUntil) throw Object.assign(new Error('Caption is already running or releasing VRAM.'), { statusCode: 409 })
+      if (!configured()) throw new Error(`Caption model is not installed. Expected ${MODEL_NAME} and ${MMPROJ_NAME} under ${models}. Run scripts/setup-caption.ps1.`)
+      if (!leaseValid(lease)) throw new Error('GPU reservation expired.')
+
+      let imagePath = request.imagePath
+      if (imagePath) {
+        if (!existsSync(imagePath)) throw new Error('imagePath does not exist on the GPU host.')
+        mkdirSync(dir(request.id), { recursive: true })
+        const dest = join(dir(request.id), `input${extname(imagePath) || '.png'}`)
+        copyFileSync(imagePath, dest)
+        imagePath = dest
+      } else {
+        imagePath = join(dir(request.id), 'input.png')
+        if (!existsSync(imagePath)) {
+          const alt = readdirSync(dir(request.id)).find(name => /^input\./i.test(name))
+          if (!alt) throw new Error('Upload an image first or pass imagePath.')
+          imagePath = join(dir(request.id), alt)
+        }
+      }
+
+      const job = {
+        id: request.id,
+        status: 'starting',
+        message: 'Preparing GPU for caption',
+        progress: 0,
+        style: request.style,
+        startedAt: Date.now(),
+        checkedAt: Date.now()
+      }
+      active = { job, child: null, cancelled: false, abort: new AbortController() }
+      const run = active
+      mkdirSync(dir(job.id), { recursive: true })
+      writeFileSync(join(dir(job.id), 'request.json'), JSON.stringify({ ...request, imagePath, prompt: captionStylePrompt(request.style) }))
+      persist(job)
+
+      try {
+        if (prepare) await prepare()
+        if (run.cancelled || !leaseValid(lease)) throw new Error('Caption start cancelled or GPU reservation expired.')
+
+        job.status = 'running'
+        job.message = 'Loading Qwen2.5-VL caption model'
+        job.progress = 5
+        persist(job)
+
+        const port = await freePort()
+        const args = [
+          '-m', modelPath(),
+          '--mmproj', mmprojPath(),
+          '--host', '127.0.0.1',
+          '--port', String(port),
+          '-ngl', String(process.env.CAPTION_NGL || '99'),
+          '-c', String(process.env.CAPTION_CTX || '4096'),
+          '--jinja'
+        ]
+        const child = spawnProcess(executable, args, {
+          windowsHide: true,
+          shell: false,
+          stdio: ['ignore', 'pipe', 'pipe'],
+          env: { ...process.env }
+        })
+        run.child = child
+        const log = chunk => appendFileSync(join(dir(job.id), 'worker.log'), chunk)
+        child.stdout.on('data', log)
+        child.stderr.on('data', log)
+
+        const watchdog = setInterval(() => {
+          if (!leaseValid(lease)) {
+            job.error = 'GPU reservation expired; caption stopped.'
+            run.abort.abort()
+            void killTree(child)
+          }
+        }, 5000)
+
+        try {
+          await waitForServer(port, run.abort.signal, 120_000, fetchImpl)
+          if (run.cancelled) throw new Error('Cancelled')
+          job.message = 'Captioning'
+          job.progress = 40
+          persist(job)
+
+          const bytes = readFileSync(imagePath)
+          const dataUrl = `data:${mimeFor(imagePath)};base64,${bytes.toString('base64')}`
+          const timeout = AbortSignal.timeout(Number(process.env.CAPTION_TIMEOUT_MS || 180_000))
+          const onAbort = () => { /* combined via race below */ }
+          run.abort.signal.addEventListener('abort', onAbort, { once: true })
+          let response
+          try {
+            response = await fetchImpl(`http://127.0.0.1:${port}/v1/chat/completions`, {
+              method: 'POST',
+              headers: { 'Content-Type': 'application/json' },
+              signal: run.abort.signal.aborted ? run.abort.signal : timeout,
+              body: JSON.stringify({
+                temperature: 0.2,
+                max_tokens: 512,
+                messages: [{
+                  role: 'user',
+                  content: [
+                    { type: 'text', text: captionStylePrompt(request.style) },
+                    { type: 'image_url', image_url: { url: dataUrl } }
+                  ]
+                }]
+              })
+            })
+          } finally {
+            run.abort.signal.removeEventListener('abort', onAbort)
+          }
+          if (run.abort.signal.aborted) throw new Error('Cancelled')
+          if (!response.ok) {
+            const detail = await response.text().catch(() => '')
+            throw new Error(`llama-server caption failed (${response.status}): ${detail.slice(0, 400)}`)
+          }
+          const payload = await response.json()
+          const text = String(payload?.choices?.[0]?.message?.content || '').trim()
+          if (!text) throw new Error('Caption model returned empty text.')
+          job.text = text
+          job.status = run.cancelled ? 'cancelled' : 'complete'
+          job.message = job.status === 'complete' ? 'Caption ready' : 'Cancelled'
+          job.progress = 100
+          job.checkedAt = Date.now()
+          persist(job)
+          return { ...job }
+        } finally {
+          clearInterval(watchdog)
+          await killTree(child)
+          // keep_alive 0: process exit unloads GGUF+mmproj from VRAM
+          holdUntil = Date.now() + 2000
+          if (active === run) active = null
+        }
+      } catch (error) {
+        job.status = run.cancelled ? 'cancelled' : 'error'
+        job.error = error.message || String(error)
+        job.message = job.error
+        job.checkedAt = Date.now()
+        persist(job)
+        if (run.child) await killTree(run.child)
+        if (active === run) active = null
+        holdUntil = Date.now() + 2000
+        throw error
+      }
+    },
+    async cancel(id) {
+      if (active?.job.id !== id) return read(id)
+      const run = active
+      run.cancelled = true
+      run.job.status = 'cancelling'
+      run.job.message = 'Cancelling'
+      persist(run.job)
+      run.abort.abort()
+      if (run.child) await killTree(run.child)
+      return read(id)
+    },
+    /** Sync one-shot used by POST /caption { imagePath, style }. */
+    async captionOnce(body, lease) {
+      const id = body.id || `caption-${Date.now()}-${Math.random().toString(36).slice(2, 10)}`
+      const result = await this.start({ ...body, id }, lease)
+      if (result.status === 'complete') return { text: result.text, id: result.id, style: result.style }
+      throw new Error(result.error || 'Caption failed.')
+    }
+  }
+}
+
+export const captionModelFiles = { model: MODEL_NAME, mmproj: MMPROJ_NAME }

scripts/comfy-host-agent.mjs

@@ -3,8 +3,8 @@ import { purgeStudio2Files } from './studio2-purge.mjs'
 import { stableMemoryArgs } from './comfy-memory-policy.mjs'
 import { createGpuReservation } from './gpu-reservation.mjs'
 import { createGpuProxy } from './gpu-proxy.mjs'
-import { createYueGpHost } from './yuegp-host.mjs'
 import { createYue2Host } from './yue2-host.mjs'
+import { createCaptionHost } from './caption-host.mjs'
 import http from 'node:http'
 import net from 'node:net'
 import { execFile, spawn } from 'node:child_process'
@@ -148,7 +148,7 @@ let proxyTarget = 0
 
 function ensureProxyListening() {
   if (proxyServer) return
-  proxyServer = createGpuProxy({ target: () => proxyTarget, reservation: gpuReservation, authorized, markWork, externalBusy: () => yueGp.busy() || yue2.busy() || upscale.busy() })
+  proxyServer = createGpuProxy({ target: () => proxyTarget, reservation: gpuReservation, authorized, markWork, externalBusy: () => yue2.busy() || upscale.busy() || caption.busy() })
   proxyServer.on('error', (error) => {
     console.log(JSON.stringify({ src: 'comfy-host-agent', event: 'proxy-error', error: String(error.message || error) }))
   })
@@ -820,7 +820,7 @@ function purgeDesktopFiles(body) {
 }
 
 const gpuReservation = createGpuReservation({ idle: async () => {
-  if (yueGp.busy() || yue2.busy() || upscale.busy()) return false
+  if (yue2.busy() || upscale.busy() || caption.busy()) return false
   if ((await trainingLock()).busy) return false
   const healthy = await syncProxy()
   if (healthy) {
@@ -831,43 +831,67 @@ const gpuReservation = createGpuReservation({ idle: async () => {
   return !(await processUp()) && !(await pythonMainUp().catch(() => true))
 } })
 
-const yueGp = createYueGpHost({
-  leaseValid: lease => gpuReservation.isOwner(lease),
-  prepare: async () => {
+async function prepareExclusiveGpu(label) {
   if ((await trainingLock()).busy) throw new Error('GPU is busy with training.')
-    if (yue2.busy()) throw new Error('YuE2 is using the GPU.')
+  if (label !== 'YuE2' && yue2.busy()) throw new Error('YuE2 is using the GPU.')
+  if (label !== 'caption' && caption.busy()) throw new Error('Caption is using the GPU.')
+  if (upscale.busy()) throw new Error('Local upscale is using the GPU.')
   const healthy = await syncProxy()
   if (healthy) {
     const queue = await fetchLocalQueue(healthy)
-      if (!queue.ok || queue.running || queue.pending) throw new Error('Comfy is busy; YuEGP cannot start.')
+    if (!queue.ok || queue.running || queue.pending) throw new Error(`Comfy is busy; ${label} cannot start.`)
+  }
+  // Stop Comfy and refuse to launch while its Python still owns VRAM.
+  if (healthy || await processUp() || await pythonMainUp().catch(() => false)) {
     await stopComfyProcesses()
     markAsleep()
   }
   if (await pythonMainUp()) throw new Error('Comfy has not stopped; retry after the GPU is free.')
 }
-})
+
+const upscale = createUpscaleHost({ leaseValid: token => gpuReservation.isOwner(token) })
 
 const yue2 = createYue2Host({
   leaseValid: lease => gpuReservation.isOwner(lease),
-  prepare: async () => {
-    if ((await trainingLock()).busy) throw new Error('GPU is busy with training.')
-    if (yueGp.busy()) throw new Error('YuEGP is using the GPU.')
-    const healthy = await syncProxy()
-    if (healthy) {
-      const queue = await fetchLocalQueue(healthy)
-      if (!queue.ok || queue.running || queue.pending) throw new Error('Comfy is busy; YuE2 cannot start.')
-      await stopComfyProcesses()
-      markAsleep()
-    }
-    if (await pythonMainUp()) throw new Error('Comfy has not stopped; retry after the GPU is free.')
-  }
+  prepare: async () => prepareExclusiveGpu('YuE2')
 })
 
-const upscale = createUpscaleHost({ leaseValid: token => gpuReservation.isOwner(token) })
+const caption = createCaptionHost({
+  leaseValid: lease => gpuReservation.isOwner(lease),
+  prepare: async () => prepareExclusiveGpu('caption')
+})
 
 async function handleControl(req, res) {
   if (!authorized(req)) return json(res, 401, { ok: false, error: 'unauthorized' })
   const url = new URL(req.url || '/', 'http://localhost')
+  if (url.pathname.startsWith('/caption')) {
+    const match = url.pathname.match(/^\/caption\/jobs\/([a-zA-Z0-9-]{12,80})(\/input|\/cancel)?$/)
+    if (req.method === 'GET' && url.pathname === '/caption/status') {
+      return json(res, 200, { configured: caption.configured(), busy: caption.busy(), backend: 'llama.cpp', modelsDir: caption.modelsDir() })
+    }
+    // Sync one-shot: POST /caption { imagePath, style } → load → caption → unload → { text }
+    if (req.method === 'POST' && url.pathname === '/caption') {
+      try {
+        const body = await readJson(req)
+        return json(res, 200, await caption.captionOnce(body, String(req.headers['x-aigen-gpu-lease'] || '')))
+      } catch (error) {
+        return json(res, error.statusCode || 400, { error: error.message || 'Caption failed' })
+      }
+    }
+    if (req.method === 'POST' && url.pathname === '/caption/jobs') {
+      return json(res, 200, await caption.start(await readJson(req), String(req.headers['x-aigen-gpu-lease'] || '')))
+    }
+    if (match && req.method === 'PUT' && match[2] === '/input') {
+      await caption.upload(match[1], req)
+      return json(res, 200, { uploaded: true })
+    }
+    if (match && req.method === 'POST' && match[2] === '/cancel') return json(res, 200, await caption.cancel(match[1]))
+    if (match && req.method === 'GET' && !match[2]) {
+      const job = caption.read(match[1])
+      return json(res, job ? 200 : 404, job || { error: 'Caption job not found' })
+    }
+    return json(res, 404, { error: 'Unknown caption endpoint' })
+  }
   if (url.pathname.startsWith('/upscale/')) {
     const match = url.pathname.match(/^\/upscale\/jobs\/([a-zA-Z0-9-]{12,80})(\/input|\/video|\/cancel)?$/)
     if (req.method === 'POST' && url.pathname === '/upscale/jobs') return json(res, 200, upscale.start(await readJson(req), String(req.headers['x-aigen-gpu-lease'] || '')))
@@ -877,23 +901,7 @@ async function handleControl(req, res) {
     if (match && req.method === 'GET' && !match[2]) { const job = upscale.read(match[1]); return json(res, job ? 200 : 404, job || { error: 'Upscale job not found' }) }
     return json(res, 404, { error: 'Unknown upscale endpoint' })
   }
-  if (url.pathname.startsWith('/yuegp/')) {
-    const match = url.pathname.match(/^\/yuegp\/jobs\/([a-zA-Z0-9-]{12,80})(\/audio|\/cancel)?$/)
-    if (req.method === 'GET' && url.pathname === '/yuegp/status') return json(res, 200, { configured: yueGp.configured(), busy: yueGp.busy(), backend: 'yuegp' })
-    if (req.method === 'POST' && url.pathname === '/yuegp/jobs') return json(res, 200, await yueGp.start(await readJson(req), String(req.headers['x-aigen-gpu-lease'] || '')))
-    if (match && req.method === 'POST' && match[2] === '/cancel') return json(res, 200, await yueGp.cancel(match[1]))
-    if (match && req.method === 'GET' && match[2] === '/audio') {
-      const path = yueGp.audio(match[1])
-      return path ? streamFile(res, path) : json(res, 404, { error: 'Audio not ready' })
-    }
-    if (match && req.method === 'GET' && !match[2]) {
-      const job = yueGp.read(match[1])
-      return json(res, job ? 200 : 404, job || { error: 'YuEGP job not found' })
-    }
-    return json(res, 404, { error: 'Unknown YuEGP endpoint' })
-  }
   if (url.pathname.startsWith('/yue2/')) {
-    if (yueGp.busy()) return json(res, 409, { message: 'YuEGP is using the GPU.' })
     const match = url.pathname.match(/^\/yue2\/jobs\/([a-zA-Z0-9-]{12,80})(\/audio|\/cancel)?$/)
     if (req.method === 'GET' && url.pathname === '/yue2/status') return json(res, 200, { configured: yue2.configured(), busy: yue2.busy(), backend: 'yue2' })
     if (req.method === 'POST' && url.pathname === '/yue2/jobs') return json(res, 200, await yue2.start(await readJson(req), String(req.headers['x-aigen-gpu-lease'] || '')))
@@ -939,15 +947,15 @@ async function handleControl(req, res) {
       proxyPort,
       gpu: gpuReservation.availability(),
       training: { busy: lastTraining.busy },
-      yuegp: { busy: yueGp.busy(), configured: yueGp.configured() },
       yue2: { busy: yue2.busy(), configured: yue2.configured() },
+      caption: { busy: caption.busy(), configured: caption.configured() },
       upscale: { busy: upscale.busy(), engine: 'realesrgan-rife', local: true }
     })
   }
   if (req.method === 'POST' && url.pathname === '/start') {
     if (upscale.busy()) return json(res, 409, { message: 'Local upscale is using the GPU.' })
-    if (yueGp.busy()) return json(res, 409, { message: 'YuEGP is using the GPU.' })
     if (yue2.busy()) return json(res, 409, { message: 'YuE2 is using the GPU.' })
+    if (caption.busy()) return json(res, 409, { message: 'Caption is using the GPU.' })
     const training = await trainingLock()
     if (training.busy) {
       return json(res, 409, {
@@ -1049,7 +1057,7 @@ const server = http.createServer(async (req, res) => {
     } else await handleControl(req, res)
   } catch (error) {
     req.resume()
-    if ((String(req.url || '').startsWith('/yuegp/') || String(req.url || '').startsWith('/yue2/') || String(req.url || '').startsWith('/upscale/')) && !res.headersSent) return json(res, error.statusCode || 400, { error: error.message || 'Music host request failed' })
+    if ((String(req.url || '').startsWith('/yue2/') || String(req.url || '').startsWith('/upscale/') || String(req.url || '').startsWith('/caption')) && !res.headersSent) return json(res, error.statusCode || 400, { error: error.message || 'Host request failed' })
     if (!res.headersSent) json(res, error.statusCode || 400, { ok: false, message: error.statusCode === 409 ? 'GPU is in use. Waiting for availability.' : 'GPU coordination request failed.' })
   }
 })

scripts/install-comfy-host-agent-launchers.ps1

@@ -1,26 +1,48 @@
-# Installs Desktop + Start Menu shortcuts that start the Comfy host agent
-# without Cursor. The scheduled task still starts the agent at logon.
+# Installs Desktop, Start Menu, and Windows Startup shortcuts that start THIS
+# clone's Comfy host agent (YuE2) and replace any stale process on 8199.
 $ErrorActionPreference = 'Stop'
 $root = (Resolve-Path (Join-Path $PSScriptRoot '..')).Path
 $vbs = Join-Path $root 'scripts\launch-comfy-host-agent.vbs'
+$startPs1 = Join-Path $root 'scripts\start-comfy-host-agent.ps1'
 if (-not (Test-Path $vbs)) { throw "Missing $vbs" }
+if (-not (Test-Path $startPs1)) { throw "Missing $startPs1" }
 
 $shell = New-Object -ComObject WScript.Shell
 $desktop = [Environment]::GetFolderPath('Desktop')
 $programs = Join-Path ([Environment]::GetFolderPath('StartMenu')) 'Programs\AIGen'
+$startup = [Environment]::GetFolderPath('Startup')
 New-Item -ItemType Directory -Force -Path $programs | Out-Null
 
-function Save-Shortcut([string]$dir) {
-  $path = Join-Path $dir 'Start Comfy Host Agent.lnk'
+function Save-Shortcut([string]$dir, [string]$name = 'Start Comfy Host Agent.lnk') {
+  $path = Join-Path $dir $name
   $lnk = $shell.CreateShortcut($path)
   $lnk.TargetPath = (Join-Path $env:SystemRoot 'System32\wscript.exe')
   $lnk.Arguments = "//B `"$vbs`""
   $lnk.WorkingDirectory = $root
   $lnk.WindowStyle = 7
-  $lnk.Description = 'Start the AIGen Comfy host agent in the background'
+  $lnk.Description = 'Start THIS AIGen Comfy host agent (YuE2) and replace any stale agent on 8199'
   $lnk.Save()
   $path
 }
 
-Save-Shortcut $desktop
-Save-Shortcut $programs
+$saved = @(
+  (Save-Shortcut $desktop),
+  (Save-Shortcut $programs),
+  (Save-Shortcut $startup 'AIGen Comfy Host Agent.lnk')
+)
+
+# Point the scheduled task at this clone when it already exists.
+$taskName = 'AIGen Comfy Host Agent'
+$existing = schtasks /Query /TN $taskName 2>$null
+if ($LASTEXITCODE -eq 0) {
+  $hidden = Join-Path $root 'scripts\start-comfy-host-agent-hidden.vbs'
+  $tr = "wscript.exe //B `"$hidden`""
+  schtasks /Change /TN $taskName /TR $tr | Out-Null
+  Write-Host "Updated scheduled task '$taskName' -> $hidden"
+}
+
+Write-Host "Installed launchers for $root"
+$saved | ForEach-Object { Write-Host "  $_" }
+Write-Host ''
+Write-Host 'One click: double-click Desktop "Start Comfy Host Agent", or log off/on so Startup runs it.'
+Write-Host 'That replaces any old node on 8199 with this clone (YuE2 /yue2/status).'

scripts/launch-comfy-host-agent.vbs

@@ -1,35 +1,14 @@
-' Start the Comfy host agent with no window. Safe to run when Cursor is closed.
-' Uses the "AIGen Comfy Host Agent" scheduled task when it exists.
+' Start THIS clone's Comfy host agent (YuE2). Replaces any stale agent on 8199.
+' Safe to run at logon / from Startup / Desktop shortcut when Cursor is closed.
 Option Explicit
-Dim sh, fso, root, node, agent, rc
+Dim sh, fso, root, ps1
 Set sh = CreateObject("WScript.Shell")
 Set fso = CreateObject("Scripting.FileSystemObject")
 
-If AgentRunning() Then WScript.Quit 0
-
-On Error Resume Next
-rc = sh.Run("schtasks /run /tn ""AIGen Comfy Host Agent""", 0, True)
-On Error GoTo 0
-WScript.Sleep 1200
-If AgentRunning() Then WScript.Quit 0
-
 root = fso.GetParentFolderName(WScript.ScriptFullName)
 root = fso.GetParentFolderName(root)
-node = """C:\Program Files\nodejs\node.exe"""
-agent = """" & root & "\scripts\comfy-host-agent.mjs"""
-If Not fso.FileExists(Replace(node, """", "")) Then WScript.Quit 1
-sh.CurrentDirectory = root
-sh.Run node & " " & agent, 0, False
+ps1 = root & "\scripts\start-comfy-host-agent.ps1"
+If Not fso.FileExists(ps1) Then WScript.Quit 1
 
-Function AgentRunning()
-  Dim svc, query, item
-  AgentRunning = False
-  Set svc = GetObject("winmgmts:\\.\root\cimv2")
-  Set query = svc.ExecQuery("SELECT CommandLine FROM Win32_Process WHERE Name='node.exe'")
-  For Each item In query
-    If InStr(LCase("" & item.CommandLine), "comfy-host-agent") > 0 Then
-      AgentRunning = True
-      Exit Function
-    End If
-  Next
-End Function
+sh.CurrentDirectory = root
+sh.Run "powershell.exe -NoProfile -WindowStyle Hidden -ExecutionPolicy Bypass -File """ & ps1 & """", 0, False

scripts/setup-caption.ps1

@@ -0,0 +1,129 @@
+# Download Qwen2.5-VL NSFW Caption V4 GGUF (Q5_K_M + mmproj only) for Studio 2 caption jobs.
+# Does NOT pull the full safetensors repo (~16.6GB).
+#
+# Example:
+#   powershell -ExecutionPolicy Bypass -File scripts\setup-caption.ps1
+#   powershell -File scripts\setup-caption.ps1 -SkipDownload
+
+[CmdletBinding()]
+param(
+  [string]$ModelsRoot = '',
+  [string]$LlamaServer = '',
+  [switch]$SkipDownload,
+  [switch]$Force
+)
+
+$ErrorActionPreference = 'Stop'
+$ProgressPreference = 'Continue'
+
+function Write-Step([string]$Message) {
+  Write-Host ""
+  Write-Host "==> $Message" -ForegroundColor Cyan
+}
+
+function Write-Ok([string]$Message) {
+  Write-Host "    $Message" -ForegroundColor Green
+}
+
+function Write-Warn([string]$Message) {
+  Write-Host "    warn  $Message" -ForegroundColor Yellow
+}
+
+function Resolve-ModelsRoot([string]$Hint) {
+  $shared = Join-Path $env:LOCALAPPDATA 'Comfy-Desktop\ComfyUI-Shared\models'
+  $candidates = @(
+    $Hint,
+    $env:COMFY_MODELS_ROOT,
+    $shared
+  ) | Where-Object { $_ } | ForEach-Object { $_.TrimEnd('\', '/') }
+
+  foreach ($path in $candidates) {
+    if (Test-Path $path) { return (Resolve-Path $path).Path }
+  }
+  New-Item -ItemType Directory -Force -Path $shared | Out-Null
+  return (Resolve-Path $shared).Path
+}
+
+function Find-HfCli {
+  foreach ($name in @('hf', 'huggingface-cli')) {
+    $cmd = Get-Command $name -ErrorAction SilentlyContinue
+    if ($cmd) { return @{ Name = $name; Source = $cmd.Source } }
+  }
+  throw "Hugging Face CLI not found (hf / huggingface-cli). Install with: pip install -U huggingface_hub"
+}
+
+function Ensure-Dir([string]$Path) {
+  if (-not (Test-Path $Path)) {
+    New-Item -ItemType Directory -Force -Path $Path | Out-Null
+  }
+}
+
+Write-Step 'Resolve models root'
+$modelsRoot = Resolve-ModelsRoot $ModelsRoot
+$captionDir = Join-Path $modelsRoot 'caption\qwen25vl-7b-nsfw-v4'
+Ensure-Dir $captionDir
+Write-Ok "Models root: $modelsRoot"
+Write-Ok "Caption dir: $captionDir"
+
+$model = 'Qwen2.5-VL-7B-NSFW-Caption-V4.Q5_K_M.gguf'
+$mmproj = 'Qwen2.5-VL-7B-NSFW-Caption-V4.mmproj-f16.gguf'
+$modelPath = Join-Path $captionDir $model
+$mmprojPath = Join-Path $captionDir $mmproj
+
+if (-not $SkipDownload) {
+  Write-Step 'Download GGUF + mmproj only (~7GB)'
+  $needModel = $Force -or -not (Test-Path $modelPath)
+  $needMmproj = $Force -or -not (Test-Path $mmprojPath)
+  if (-not $needModel -and -not $needMmproj) {
+    Write-Ok 'Both weight files already present.'
+  } else {
+    $hf = Find-HfCli
+    if ($hf.Name -eq 'hf') {
+      & $hf.Source download mradermacher/Qwen2.5-VL-7B-NSFW-Caption-V4-GGUF `
+        --include $model `
+        --include $mmproj `
+        --local-dir $captionDir
+    } else {
+      & $hf.Source download mradermacher/Qwen2.5-VL-7B-NSFW-Caption-V4-GGUF `
+        --include $model `
+        --include $mmproj `
+        --local-dir $captionDir
+    }
+    if (-not (Test-Path $modelPath)) { throw "Missing $model after download." }
+    if (-not (Test-Path $mmprojPath)) { throw "Missing $mmproj after download." }
+    Write-Ok 'Download complete.'
+  }
+} else {
+  Write-Warn 'SkipDownload set — not fetching weights.'
+}
+
+Write-Step 'Locate llama-server'
+$server = $LlamaServer
+if (-not $server) { $server = $env:CAPTION_LLAMA_SERVER }
+if (-not $server) {
+  $cmd = Get-Command llama-server -ErrorAction SilentlyContinue
+  if ($cmd) { $server = $cmd.Source }
+}
+if ($server) {
+  Write-Ok "llama-server: $server"
+} else {
+  Write-Warn 'llama-server not on PATH. Install with: winget install ggml.llamacpp'
+  Write-Warn 'Or set CAPTION_LLAMA_SERVER to a CUDA build of llama-server.exe'
+}
+
+$ready = @{
+  model = $modelPath
+  mmproj = $mmprojPath
+  modelsDir = $captionDir
+  llamaServer = $server
+  backend = 'llama.cpp'
+  updatedAt = (Get-Date).ToString('o')
+}
+$readyPath = Join-Path $captionDir 'aigen-ready.json'
+$ready | ConvertTo-Json | Set-Content -Path $readyPath -Encoding utf8
+Write-Ok "Wrote $readyPath"
+Write-Host ""
+Write-Host "Caption host looks for:" -ForegroundColor Cyan
+Write-Host "  $modelPath"
+Write-Host "  $mmprojPath"
+Write-Host "Restart the Comfy host agent after install so /caption/status reports configured=true."

scripts/setup-qwen21.ps1

@@ -0,0 +1,225 @@
+# Download Qwen Image 2.1 GGUF + companions into the live Comfy Desktop Shared models tree.
+# Also ensures city96/ComfyUI-GGUF is present on the ComfyUI (1) install.
+#
+# Example:
+#   powershell -ExecutionPolicy Bypass -File scripts\setup-qwen21.ps1
+#   powershell -File scripts\setup-qwen21.ps1 -SkipDownload
+#   powershell -File scripts\setup-qwen21.ps1 -ComfyRoot "C:\...\ComfyUI (1)\ComfyUI"
+
+[CmdletBinding()]
+param(
+  [string]$ComfyRoot = '',
+  [string]$ModelsRoot = '',
+  [switch]$SkipDownload,
+  [switch]$SkipNode,
+  [switch]$Force
+)
+
+$ErrorActionPreference = 'Stop'
+$ProgressPreference = 'Continue'
+
+function Write-Step([string]$Message) {
+  Write-Host ""
+  Write-Host "==> $Message" -ForegroundColor Cyan
+}
+
+function Write-Ok([string]$Message) {
+  Write-Host "    $Message" -ForegroundColor Green
+}
+
+function Write-Warn([string]$Message) {
+  Write-Host "    warn  $Message" -ForegroundColor Yellow
+}
+
+function Resolve-ComfyRoot([string]$Hint) {
+  $candidates = @(
+    $Hint,
+    $env:COMFY_ROOT,
+    $env:COMFYUI_ROOT,
+    (Join-Path $env:LOCALAPPDATA 'Comfy-Desktop\ComfyUI-Installs\ComfyUI (1)\ComfyUI'),
+    (Join-Path $env:LOCALAPPDATA 'Comfy-Desktop\ComfyUI-Installs\ComfyUI\ComfyUI'),
+    'C:\ComfyUI',
+    (Join-Path $env:USERPROFILE 'ComfyUI')
+  ) | Where-Object { $_ } | ForEach-Object { $_.TrimEnd('\', '/') }
+
+  foreach ($path in $candidates) {
+    if (Test-Path (Join-Path $path 'main.py')) { return (Resolve-Path $path).Path }
+    if (Test-Path (Join-Path $path 'ComfyUI\main.py')) { return (Resolve-Path (Join-Path $path 'ComfyUI')).Path }
+  }
+  throw "Could not find ComfyUI root (main.py). Pass -ComfyRoot."
+}
+
+function Resolve-ModelsRoot([string]$Hint, [string]$ComfyRootPath) {
+  $shared = Join-Path $env:LOCALAPPDATA 'Comfy-Desktop\ComfyUI-Shared\models'
+  $candidates = @(
+    $Hint,
+    $env:COMFY_MODELS_ROOT,
+    $shared,
+    (Join-Path $ComfyRootPath 'models')
+  ) | Where-Object { $_ } | ForEach-Object { $_.TrimEnd('\', '/') }
+
+  foreach ($path in $candidates) {
+    if (Test-Path $path) { return (Resolve-Path $path).Path }
+  }
+  New-Item -ItemType Directory -Force -Path $shared | Out-Null
+  return (Resolve-Path $shared).Path
+}
+
+function Find-HfCli {
+  foreach ($name in @('hf', 'huggingface-cli')) {
+    $cmd = Get-Command $name -ErrorAction SilentlyContinue
+    if ($cmd) { return $cmd.Source }
+  }
+  throw "Hugging Face CLI not found (hf / huggingface-cli). Install with: pip install -U huggingface_hub"
+}
+
+function Ensure-Dir([string]$Path) {
+  if (-not (Test-Path $Path)) {
+    New-Item -ItemType Directory -Force -Path $Path | Out-Null
+  }
+}
+
+function Assert-ModelFile([string]$Path, [long]$MinBytes) {
+  if (-not (Test-Path $Path)) { throw "Missing model file: $Path" }
+  $item = Get-Item $Path
+  if ($item.Length -lt $MinBytes) {
+    throw "Suspicious size for $Path ($($item.Length) bytes). Expected >= $MinBytes. Delete and re-run."
+  }
+  Write-Ok ("{0}  {1:N0} bytes" -f $item.FullName, $item.Length)
+}
+
+function Flatten-IfNested([string]$DestDir, [string]$FileName) {
+  $nested = Get-ChildItem $DestDir -Recurse -File -Filter $FileName -ErrorAction SilentlyContinue |
+    Where-Object { $_.DirectoryName -ne $DestDir } |
+    Select-Object -First 1
+  if ($nested) {
+    $target = Join-Path $DestDir $FileName
+    if ((Test-Path $target) -and -not $Force) {
+      Write-Warn "Nested copy found but destination already exists: $target"
+      return
+    }
+    Move-Item -Force $nested.FullName $target
+    Write-Ok "Flattened $($nested.FullName) -> $target"
+  }
+}
+
+function Invoke-HfFile {
+  param(
+    [string]$Cli,
+    [string]$Repo,
+    [string]$File,
+    [string]$LocalDir
+  )
+  Ensure-Dir $LocalDir
+  $dest = Join-Path $LocalDir (Split-Path $File -Leaf)
+  if ((Test-Path $dest) -and -not $Force) {
+    $len = (Get-Item $dest).Length
+    if ($len -gt 1MB) {
+      Write-Ok "Already present: $dest ($len bytes)"
+      return
+    }
+  }
+  Write-Step "hf download $Repo $File"
+  & $Cli download $Repo $File --local-dir $LocalDir
+  if ($LASTEXITCODE -ne 0) { throw "hf download failed for $Repo / $File (exit $LASTEXITCODE)" }
+  Flatten-IfNested -DestDir $LocalDir -FileName (Split-Path $File -Leaf)
+}
+
+# --- resolve paths ---
+Write-Step "Resolve ComfyUI root"
+$root = Resolve-ComfyRoot $ComfyRoot
+Write-Ok $root
+
+Write-Step "Resolve models root (Desktop Shared preferred)"
+$models = Resolve-ModelsRoot $ModelsRoot $root
+Write-Ok $models
+
+$Diff = Join-Path $models 'diffusion_models'
+$TE   = Join-Path $models 'text_encoders'
+$VAE  = Join-Path $models 'vae'
+Ensure-Dir $Diff
+Ensure-Dir $TE
+Ensure-Dir $VAE
+
+if (-not $SkipDownload) {
+  $cli = Find-HfCli
+  Write-Ok "HF CLI: $cli"
+
+  # DiT GGUF only (no Q6/Q5/Q4, no full bf16 DiT)
+  # Exact invocation required on this host:
+  #   hf download hf://abenzerps/Qwen-Image-2.1-Uncensored-GGUF/qwen-image-2.1-Q8_0.gguf
+  # That form writes into the Hugging Face hub cache (not --local-dir).
+  Write-Step "hf download hf://abenzerps/Qwen-Image-2.1-Uncensored-GGUF/qwen-image-2.1-Q8_0.gguf"
+  $ditName = 'qwen-image-2.1-Q8_0.gguf'
+  $ditDest = Join-Path $Diff $ditName
+  if ((Test-Path $ditDest) -and -not $Force -and (Get-Item $ditDest).Length -gt 1MB) {
+    Write-Ok "Already present: $ditDest ($((Get-Item $ditDest).Length) bytes)"
+  } else {
+    & $cli download "hf://abenzerps/Qwen-Image-2.1-Uncensored-GGUF/$ditName"
+    if ($LASTEXITCODE -ne 0) { throw "hf download failed for DiT GGUF (exit $LASTEXITCODE)" }
+    $cacheRoot = Join-Path $env:USERPROFILE '.cache\huggingface\hub\models--abenzerps--Qwen-Image-2.1-Uncensored-GGUF'
+    $found = Get-ChildItem $cacheRoot -Recurse -File -Filter $ditName -ErrorAction SilentlyContinue |
+      Sort-Object LastWriteTime -Descending |
+      Select-Object -First 1
+    if (-not $found) { throw "hf download finished but $ditName was not found under $cacheRoot" }
+    Ensure-Dir $Diff
+    Copy-Item -Force $found.FullName $ditDest
+    Write-Ok "Copied $($found.FullName) -> $ditDest ($((Get-Item $ditDest).Length) bytes)"
+  }
+
+  # abenzerps ships kv_count=0 + Q8_0 1D RMSNorms (packed 136 vs logical 128).
+  # Tag architecture=qwen_image and promote small/1D tensors to F32 for ComfyUI-GGUF.
+  Write-Step "Tag DiT GGUF (architecture + F32 norms)"
+  $tagger = Join-Path $PSScriptRoot 'tag-qwen21-gguf.py'
+  $pyCandidates = @(
+    (Join-Path $root '.venv\Scripts\python.exe'),
+    (Join-Path (Split-Path $root) 'standalone-env\python.exe'),
+    (Get-Command python -ErrorAction SilentlyContinue | Select-Object -ExpandProperty Source)
+  ) | Where-Object { $_ -and (Test-Path $_) }
+  $py = $pyCandidates | Select-Object -First 1
+  if (-not $py) { throw "No Python found to run $tagger" }
+  $fixed = Join-Path $Diff 'qwen-image-2.1-Q8_0.fixed.gguf'
+  & $py $tagger $ditDest $fixed
+  if ($LASTEXITCODE -ne 0) { throw "tag-qwen21-gguf.py failed (exit $LASTEXITCODE)" }
+  Copy-Item -Force $fixed $ditDest
+  Write-Ok "Installed tagged GGUF at $ditDest"
+
+  # TE + VAE companions from Comfy-Org (INT8 TE for 16 GB VRAM)
+  Invoke-HfFile -Cli $cli -Repo 'Comfy-Org/Qwen-Image-2.1' -File 'text_encoders/qwen3vl_8b_int8_convrot.safetensors' -LocalDir $TE
+  Invoke-HfFile -Cli $cli -Repo 'Comfy-Org/Qwen-Image-2.1' -File 'vae/qwen_image_2.1_vae_bf16.safetensors' -LocalDir $VAE
+}
+
+Write-Step "Verify weights"
+Assert-ModelFile (Join-Path $Diff 'qwen-image-2.1-Q8_0.gguf') 7000000000
+Assert-ModelFile (Join-Path $TE 'qwen3vl_8b_int8_convrot.safetensors') 1000000000
+Assert-ModelFile (Join-Path $VAE 'qwen_image_2.1_vae_bf16.safetensors') 100000000
+
+if (-not $SkipNode) {
+  Write-Step "city96 ComfyUI-GGUF"
+  $nodes = Join-Path $root 'custom_nodes\ComfyUI-GGUF'
+  if (-not (Test-Path $nodes)) {
+    git clone https://github.com/city96/ComfyUI-GGUF $nodes
+    if ($LASTEXITCODE -ne 0) { throw "git clone ComfyUI-GGUF failed" }
+    Write-Ok "Cloned $nodes"
+  } else {
+    Write-Ok "Already present: $nodes"
+  }
+
+  $py = Join-Path $root '.venv\Scripts\python.exe'
+  if (-not (Test-Path $py)) {
+    $py = Join-Path (Split-Path $root) 'standalone-env\python.exe'
+  }
+  if (Test-Path $py) {
+    Write-Step "Ensure gguf package in Comfy python ($py)"
+    & $py -m pip install -q gguf
+    if ($LASTEXITCODE -ne 0) { Write-Warn "pip install gguf returned $LASTEXITCODE" }
+    else { Write-Ok "gguf importable" }
+  } else {
+    Write-Warn "Comfy python not found; install gguf into the interpreter that runs main.py"
+  }
+}
+
+Write-Host ""
+Write-Host "Done. Restart Comfy only when /status shows gpu.busy=false." -ForegroundColor Cyan
+Write-Host "ComfyRoot=$root"
+Write-Host "ModelsRoot=$models"

scripts/start-comfy-host-agent-hidden.vbs

@@ -1,11 +1,15 @@
-' Launch the Comfy host agent with no console window.
-' Window style 0 = hidden. Wait = True so Task Scheduler keeps the task marked Running.
+' Launch the Comfy host agent with no console window (Task Scheduler wait mode).
+' Replaces any stale agent on 8199, then runs THIS clone's agent until it exits.
 Option Explicit
-Dim sh, root, node, agent
+Dim sh, fso, root, ps1
 Set sh = CreateObject("WScript.Shell")
-root = CreateObject("Scripting.FileSystemObject").GetParentFolderName(WScript.ScriptFullName)
-root = CreateObject("Scripting.FileSystemObject").GetParentFolderName(root)
-node = """C:\Program Files\nodejs\node.exe"""
-agent = """" & root & "\scripts\comfy-host-agent.mjs"""
+Set fso = CreateObject("Scripting.FileSystemObject")
+
+root = fso.GetParentFolderName(WScript.ScriptFullName)
+root = fso.GetParentFolderName(root)
+ps1 = root & "\scripts\start-comfy-host-agent.ps1"
+If Not fso.FileExists(ps1) Then WScript.Quit 1
+
 sh.CurrentDirectory = root
-sh.Run node & " " & agent, 0, True
+' Wait=True so Task Scheduler keeps the task marked Running while the agent lives.
+sh.Run "powershell.exe -NoProfile -WindowStyle Hidden -ExecutionPolicy Bypass -File """ & ps1 & """ -Wait", 0, True

scripts/start-comfy-host-agent.ps1

@@ -0,0 +1,34 @@
+# Replace any stale Comfy host agent, then start THIS clone's agent on 8199.
+# Called by launch-comfy-host-agent.vbs (Startup / Desktop / Start Menu).
+# -Wait: run in-foreground for Task Scheduler (keeps task marked Running).
+param([switch]$Wait)
+$ErrorActionPreference = 'SilentlyContinue'
+$root = (Resolve-Path (Join-Path $PSScriptRoot '..')).Path
+$node = 'C:\Program Files\nodejs\node.exe'
+$agent = Join-Path $root 'scripts\comfy-host-agent.mjs'
+if (-not (Test-Path $node)) { throw "Missing $node" }
+if (-not (Test-Path $agent)) { throw "Missing $agent" }
+
+Get-CimInstance Win32_Process -Filter "Name='node.exe'" |
+  Where-Object { $_.CommandLine -match 'comfy-host-agent' } |
+  ForEach-Object { Stop-Process -Id $_.ProcessId -Force }
+
+Get-NetTCPConnection -LocalPort 8199 -State Listen -ErrorAction SilentlyContinue |
+  ForEach-Object { Stop-Process -Id $_.OwningProcess -Force }
+
+Start-Sleep -Milliseconds 400
+
+if ($Wait) {
+  Set-Location $root
+  & $node $agent
+  exit $LASTEXITCODE
+}
+
+$psi = New-Object System.Diagnostics.ProcessStartInfo
+$psi.FileName = $node
+$psi.Arguments = "`"$agent`""
+$psi.WorkingDirectory = $root
+$psi.UseShellExecute = $false
+$psi.CreateNoWindow = $true
+$psi.WindowStyle = [System.Diagnostics.ProcessWindowStyle]::Hidden
+[void][System.Diagnostics.Process]::Start($psi)

scripts/tag-qwen21-gguf.py

@@ -0,0 +1,132 @@
+#!/usr/bin/env python3
+"""Make abenzerps Qwen Image 2.1 DiT GGUF loadable in city96 ComfyUI-GGUF.
+
+abenzerps/Qwen-Image-2.1-Uncensored-GGUF ships Q8_0 with:
+  - kv_count=0 (no general.architecture)
+  - 1D RMSNorm weights quantized to Q8_0 (logical 128 -> packed 136),
+    which breaks Comfy's fused rms_rope path
+
+This rewrite:
+  1. Adds general.architecture=qwen_image
+  2. Dequantizes small / 1D tensors to F32 (city96 convert keeps them hiprec)
+"""
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+
+# Match city96 ComfyUI-GGUF/tools/convert.py QUANTIZATION_THRESHOLD
+QUANTIZATION_THRESHOLD = 1024
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("src", type=Path)
+    ap.add_argument("dst", type=Path, nargs="?", default=None)
+    ap.add_argument("--arch", default="qwen_image")
+    ap.add_argument("--name", default="qwen-image-2.1")
+    ap.add_argument("--inplace", action="store_true", help="Replace src after a successful tag")
+    args = ap.parse_args()
+
+    import gguf
+    import numpy as np
+
+    src = args.src.resolve()
+    if not src.is_file():
+        print(f"missing source: {src}", file=sys.stderr)
+        return 1
+
+    dst = (args.dst.resolve() if args.dst else src.with_name(src.stem + ".tagged.gguf"))
+
+    reader = gguf.GGUFReader(str(src))
+
+    def get_field(name: str):
+        field = reader.fields.get(name)
+        if field is None:
+            return None
+        try:
+            return field.contents()
+        except Exception:
+            return None
+
+    if dst == src:
+        dst = src.with_name(src.stem + ".retag.gguf")
+
+    f32 = gguf.GGMLQuantizationType.F32
+    compat = {gguf.GGMLQuantizationType.F32, gguf.GGMLQuantizationType.F16, gguf.GGMLQuantizationType.BF16}
+
+    print(f"rewriting {src} -> {dst} (arch={args.arch}, tensors={len(reader.tensors)})")
+    writer = gguf.GGUFWriter(str(dst), arch=args.arch, use_temp_file=True)
+    writer.add_name(args.name)
+    try:
+        writer.add_type("model")
+    except Exception:
+        pass
+
+    promoted = 0
+    for tensor in reader.tensors:
+        logical = tuple(int(x) for x in tensor.shape)
+        # GGUF stores dims reversed vs torch; logical numel is what matters
+        numel = 1
+        for d in logical:
+            numel *= d
+        qtype = tensor.tensor_type
+        data = tensor.data
+        if hasattr(data, "copy"):
+            data = data.copy()
+
+        needs_f32 = qtype not in compat and (len(logical) <= 1 or numel <= QUANTIZATION_THRESHOLD)
+        if needs_f32:
+            # Dequantize packed blocks -> float32 with logical shape (GGUF order)
+            dequant = gguf.quants.dequantize(np.asarray(data), qtype).astype(np.float32, copy=False)
+            expected = numel
+            if dequant.size != expected:
+                print(
+                    f"dequant size mismatch {tensor.name}: got {dequant.size} want {expected} "
+                    f"shape={logical} qtype={qtype}",
+                    file=sys.stderr,
+                )
+                return 3
+            data = dequant.reshape(logical)
+            qtype = f32
+            promoted += 1
+
+        writer.add_tensor(tensor.name, data, raw_dtype=qtype)
+
+    writer.write_header_to_file()
+    writer.write_kv_data_to_file()
+    writer.write_tensors_to_file(progress=True)
+    writer.close()
+
+    check = gguf.GGUFReader(str(dst))
+    field = check.fields.get("general.architecture")
+    got = field.contents() if field is not None else None
+    print(
+        f"verified architecture={got!r} tensors={len(check.tensors)} "
+        f"promoted_f32={promoted} size={dst.stat().st_size}"
+    )
+    if got != args.arch or len(check.tensors) != len(reader.tensors):
+        print("tag failed", file=sys.stderr)
+        return 2
+
+    # Spot-check a known norm weight is F32 @ 128
+    sample = next((t for t in check.tensors if t.name.endswith("attn.norm_q.weight")), None)
+    if sample is not None:
+        print(
+            f"sample {sample.name}: shape={tuple(int(x) for x in sample.shape)} "
+            f"type={sample.tensor_type.name} data_shape={tuple(sample.data.shape)}"
+        )
+
+    if args.inplace:
+        bak = src.with_suffix(src.suffix + ".untagged.bak")
+        if bak.exists():
+            bak.unlink()
+        src.replace(bak)
+        dst.replace(src)
+        print(f"inplace: {src} (backup {bak})")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/yue2-host.mjs

@@ -5,14 +5,12 @@ import { fileURLToPath } from 'node:url'
 
 export function validateYue2Request(body) {
   const duration = body.duration ?? 60
-  if (!Number.isInteger(duration) || duration < 30 || duration > 150) throw new Error('YuE2 duration must be 30–150 seconds.')
+  if (!Number.isInteger(duration) || duration < 15 || duration > 150) throw new Error('YuE2 target length must be 15–150 seconds.')
   if (!Number.isInteger(body.seed) || body.seed < 0 || body.seed > 2147483647) throw new Error('Invalid seed.')
   if (!/^[a-zA-Z0-9-]{12,80}$/.test(body.id || '')) throw new Error('Invalid job ID.')
   const tags = String(body.tags || '').trim()
   const lyrics = String(body.lyrics || '').trim()
-  if (!tags || tags.length > 2000 || !lyrics || lyrics.length > 8000) throw new Error('Genre tags and one lyric section are required.')
-  const sections = [...lyrics.matchAll(/\[([^\]]+)\]\s*([^\[]*)/gs)].filter(m => m[2].trim())
-  if (sections.length !== 1) throw new Error('YuE2 requires one non-empty lyric section.')
+  if (!tags || tags.length > 2000 || !lyrics || lyrics.length > 8000) throw new Error('Genre tags and non-empty lyrics are required.')
   return { id: body.id, duration, seed: body.seed, tags, lyrics }
 }
 

scripts/yue2-worker.py

@@ -1,6 +1,6 @@
 """Headless YuE2 adapter: plan → generate_semantic → synthesize → decode.
 
-No Comfy imports. No score editor, covers, or auto-fallback to YuEGP.
+No Comfy imports. No score editor, covers, or auto-fallback to YuEGP / ACE.
 Stdout AIGEN_EVENT lines are consumed by the device-local host agent.
 """
 import argparse
@@ -12,20 +12,71 @@ import sys
 import time
 import threading
 
+# Must be set before the first torch import on this 16GB Windows host.
+os.environ.setdefault('PYTORCH_CUDA_ALLOC_CONF', 'expandable_segments:True')
+
 
 def emit(**event):
     print('AIGEN_EVENT ' + json.dumps(event), flush=True)
 
 
 def normalize_lyrics(text):
-    # Same one-section rule as YuEGP field validation.
+    # YuE2 accepts multiple [Verse]/[Chorus] headings. Only wrap when none exist.
     import re
-    sections = re.findall(r'\[([^\]]+)\]\s*([^\[]*)', text.strip(), re.S)
-    sections = [(name, words.strip()) for name, words in sections if words.strip()]
-    if len(sections) != 1:
-        raise ValueError('YuE2 requires one non-empty lyric section. Combine the lyrics under one heading.')
-    name, words = sections[0]
-    return '[' + re.sub(r'\W+', '', name).lower() + ']\n' + words + '\n\n'
+    text = str(text or '').strip()
+    if not text:
+        raise ValueError('YuE2 requires non-empty lyrics.')
+    if re.search(r'\[[^\]]+\]', text):
+        return text if text.endswith('\n') else text + '\n'
+    return '[song]\n' + text + '\n\n'
+
+
+def is_cuda_oom(error):
+    message = str(error).lower()
+    return 'out of memory' in message or ('cuda' in message and 'alloc' in message) or 'cudnn_status_alloc_failed' in message
+
+
+def resolve_attention_backend(torch_mod):
+    """Never flash. Prefer torch-eager (no CUDA graphs) on the 16GB 5080."""
+    # Windows wheels expose flash ops without USE_FLASH_ATTENTION; never select flash.
+    # torch-eager disables GraphAR / CUDA graphs via YuE2Pipeline.backend.
+    return 'sdpa', 'torch-eager'
+
+
+def patch_graph_attention(attention_backend):
+    from yue2.cuda_graph import GraphAR
+    original = GraphAR.__init__
+
+    def init(self, model, prefixes, max_tokens, *, capture=True, attention_backend='auto', fuse_projections=False):
+        if attention_backend in ('auto', 'flash'):
+            attention_backend = patch_graph_attention.forced
+        return original(self, model, prefixes, max_tokens, capture=capture,
+                        attention_backend=attention_backend, fuse_projections=fuse_projections)
+
+    patch_graph_attention.forced = attention_backend
+    GraphAR.__init__ = init
+
+
+def free_cuda(torch_mod):
+    gc.collect()
+    if torch_mod.cuda.is_available():
+        torch_mod.cuda.empty_cache()
+        torch_mod.cuda.synchronize()
+
+
+def memory_snapshot(torch_mod, stage):
+    if not torch_mod.cuda.is_available():
+        emit(stage=stage, message=f'YuE2 {stage}', cuda=False)
+        return
+    free, total = torch_mod.cuda.mem_get_info()
+    emit(
+        stage=stage,
+        message=f'YuE2 {stage}',
+        gpuAllocatedMiB=round(torch_mod.cuda.memory_allocated() / 1024 ** 2),
+        gpuReservedMiB=round(torch_mod.cuda.memory_reserved() / 1024 ** 2),
+        gpuFreeMiB=round(free / 1024 ** 2),
+        gpuTotalMiB=round(total / 1024 ** 2),
+    )
 
 
 def main():
@@ -46,8 +97,10 @@ def main():
     root = Path(cli.root).resolve()
     output = Path(cli.request).resolve().parent
     duration = int(request.get('duration', 60))
-    if not 30 <= duration <= 150:
-        raise ValueError('Duration must be 30–150 seconds.')
+    if not 15 <= duration <= 150:
+        raise ValueError('Target length must be 15–150 seconds.')
+    # VAE downsampling_ratio 1920 @ 48 kHz → 25 semantic / latent frames per second.
+    max_tokens = max(200, min(9000, duration * 25))
     lyrics = normalize_lyrics(request['lyrics'])
     style = ' '.join(str(request['tags']).split())
     seed = int(request['seed'])
@@ -56,28 +109,71 @@ def main():
     os.chdir(root)
     if str(root) not in sys.path:
         sys.path.insert(0, str(root))
-    emit(stage='loading', progress=1, message='Loading YuE2', model=model, vae=vae, duration=duration)
     import torch
     import soundfile as sf
+    attention_backend, pipeline_backend = resolve_attention_backend(torch)
+    emit(stage='loading', progress=1, message='Loading YuE2', model=model, vae=vae,
+         duration=duration, maxTokens=max_tokens, attention_backend=attention_backend,
+         pipeline_backend=pipeline_backend,
+         cudaAllocConf=os.environ.get('PYTORCH_CUDA_ALLOC_CONF'))
     from yue2 import YuE2Pipeline
+    from yue2.protocol import Sampling
     if not torch.cuda.is_available():
         raise RuntimeError('YuE2 requires a CUDA GPU; CPU fallback is disabled.')
-    pipe_kwargs = dict(style=style, lyrics=lyrics, cot='full', seed=seed)
-    # Duration is kept for library metadata and validation. Upstream one-shot
-    # requests do not take a seconds field; song length follows the plan.
-    with YuE2Pipeline.from_pretrained(model, vae=vae, device='cuda') as pipe:
-        emit(stage='plan', message='Planning melody and chords', progress=5)
+    patch_graph_attention(attention_backend)
+    free_cuda(torch)
+    memory_snapshot(torch, 'memory-before-load')
+    # Prefer tiled VAE; do not cap the process with set_per_process_memory_fraction.
+    pipe_load = dict(
+        device='cuda',
+        backend=pipeline_backend,
+        vae_core_frames=512,
+        offload_ar=True,
+        progress=False,
+    )
+    cot = 'full'
+    pipe_kwargs = dict(style=style, lyrics=lyrics, cot=cot, seed=seed)
+    # Cap semantic length from target seconds. Do not FFmpeg-trim after decode.
+    semantic_sampling = Sampling(max_tokens=max_tokens, min_tokens=min(200, max_tokens))
+    # YuE2Pipeline.__init__ always calls set_per_process_memory_fraction; skip it on this 16GB host.
+    _set_fraction = torch.cuda.set_per_process_memory_fraction
+    torch.cuda.set_per_process_memory_fraction = lambda *args, **kwargs: None
+    try:
+        pipe_cm = YuE2Pipeline.from_pretrained(model, vae=vae, **pipe_load)
+    finally:
+        torch.cuda.set_per_process_memory_fraction = _set_fraction
+    with pipe_cm as pipe:
+        memory_snapshot(torch, 'memory-loaded')
+        emit(stage='plan', message='Planning melody and chords', progress=5, cot=cot)
+        try:
             plan = pipe.plan(**pipe_kwargs)
-        emit(stage='semantic', message='Generating semantic tokens', progress=25)
-        semantic = pipe.generate_semantic(plan)
-        emit(stage='synthesize', message='Synthesizing acoustic latents', progress=55)
+        except Exception as error:
+            if cot != 'full' or not is_cuda_oom(error):
+                raise
+            free_cuda(torch)
+            cot = 'melody'
+            pipe_kwargs['cot'] = cot
+            emit(stage='plan', message='Full CoT OOM; retrying melody CoT', progress=5, cot=cot,
+                 error=str(error))
+            memory_snapshot(torch, 'memory-before-melody-plan')
+            plan = pipe.plan(**pipe_kwargs)
+        free_cuda(torch)
+        memory_snapshot(torch, 'memory-after-plan')
+        emit(stage='semantic', message='Generating semantic tokens', progress=25, cot=cot,
+             maxTokens=max_tokens, targetSeconds=duration)
+        semantic = pipe.generate_semantic(plan, sampling=semantic_sampling)
+        free_cuda(torch)
+        memory_snapshot(torch, 'memory-after-semantic')
+        emit(stage='synthesize', message='Synthesizing acoustic latents', progress=55, cot=cot)
         latents = pipe.synthesize(semantic)
-        emit(stage='decode', message='Decoding audio', progress=80)
-        audio = pipe.decode(latents)
-    # Context exit unloads the pipeline. Clear any residual CUDA cache.
-    gc.collect()
-    if torch.cuda.is_available():
-        torch.cuda.empty_cache()
+        free_cuda(torch)
+        memory_snapshot(torch, 'memory-after-synthesize')
+        emit(stage='decode', message='Decoding audio (tiled)', progress=80, cot=cot)
+        # full=False uses decode_tiled with vae_core_frames / halo; model offloads to CPU after.
+        audio = pipe.decode(latents, full=False)
+        free_cuda(torch)
+        memory_snapshot(torch, 'memory-after-decode')
+    free_cuda(torch)
     wave = audio
     if hasattr(audio, 'detach'):
         wave = audio.detach().cpu().numpy()
@@ -93,7 +189,7 @@ def main():
     info = sf.info(str(target))
     if info.frames <= 0:
         raise RuntimeError('YuE2 produced empty audio.')
-    emit(stage='complete', message='Audio ready', progress=100, duration=info.duration)
+    emit(stage='complete', message='Audio ready', progress=100, duration=info.duration, cot=cot)
 
 
 if __name__ == '__main__':

server/api/generate/music.post.ts

@@ -1,16 +1,7 @@
-import { yueGpConfigured } from '~/server/utils/yueGp'
 import { yue2Configured } from '~/server/utils/yue2'
 import { addStudioJob, kickStudioQueue, listStudioJobs, videoJobsBusy } from '~/server/utils/studioQueue'
-import { comfyConfigured } from '~/server/utils/comfy'
 import {
-  clampLyricsStrength,
-  clampMusicCfg,
   clampMusicDuration,
-  clampMusicSteps,
-  MUSIC_CFG_DEFAULT,
-  MUSIC_CFG_DEFAULT_15,
-  MUSIC_STEPS_DEFAULT,
-  MUSIC_STEPS_DEFAULT_15,
   parseMusicEngine,
   parseMusicSeed,
   sanitizeMusicLyrics,
@@ -23,37 +14,28 @@ export default defineEventHandler(async (event) => {
     tags?: unknown
     lyrics?: unknown
     duration?: unknown
-    steps?: unknown
     seed?: unknown
-    cfg?: unknown
-    lyricsStrength?: unknown
-    instrumental?: unknown
     folderId?: unknown
     name?: unknown
-    yueProfile?: unknown
     engine?: unknown
-    sampler_name?: unknown
-    scheduler?: unknown
   }>(event).catch(() => ({} as Record<string, unknown>))
 
   const engine = parseMusicEngine(body.engine)
-  if (engine === 'yue' && !yueGpConfigured()) throw createError({ statusCode: 503, statusMessage: 'YuEGP host is not configured. Set COMFY_CONTROL_URL.' })
-  if (engine === 'yue2' && !yue2Configured()) throw createError({ statusCode: 503, statusMessage: 'YuE2 host is not configured. Set COMFY_CONTROL_URL.' })
-  if (engine !== 'yue' && engine !== 'yue2' && !comfyConfigured()) {
-    throw createError({
-      statusCode: 503,
-      statusMessage: 'Beast ComfyUI is not configured. Set COMFY_HOST.'
-    })
+  if (engine !== 'yue2') {
+    throw createError({ statusCode: 400, statusMessage: 'Only YuE2 is supported for music generation.' })
+  }
+  if (!yue2Configured()) {
+    throw createError({ statusCode: 503, statusMessage: 'YuE2 host is not configured. Set COMFY_CONTROL_URL.' })
   }
 
-  const instrumental = body.instrumental === true || body.instrumental === 'true'
   const tags = sanitizeMusicTags(body.tags)
-  const lyrics = sanitizeMusicLyrics(body.lyrics, instrumental)
+  const lyrics = sanitizeMusicLyrics(body.lyrics, false)
   if (!tags) {
     throw createError({ statusCode: 400, statusMessage: 'Genre and style tags are required' })
   }
-  if (!instrumental && !lyrics) {
-    throw createError({ statusCode: 400, statusMessage: 'Lyrics are required unless Instrumental is on' })
+  const lyricsProblem = yueLyricsProblem(lyrics)
+  if (lyricsProblem) {
+    throw createError({ statusCode: 400, statusMessage: lyricsProblem })
   }
 
   const ownerKey = libraryOwnerKey(event)
@@ -68,31 +50,12 @@ export default defineEventHandler(async (event) => {
   const destFolder = library.folders.find(folder => folder.id === folderId)
   const folderLocked = Boolean(destFolder?.protected && !destFolder.unlocked)
 
-  if ((engine === 'yue' || engine === 'yue2') && instrumental) throw createError({ statusCode: 400, statusMessage: 'YuE currently requires lyrics. Use ACE for instrumental music.' })
-  const yueProfile = body.yueProfile ?? 1
-  if (engine === 'yue' && yueProfile !== 1 && yueProfile !== 3) throw createError({ statusCode: 400, statusMessage: 'Choose YuEGP profile 1 or manual fallback 3.' })
-  if ((engine === 'yue' || engine === 'yue2') && !instrumental) {
-    const problem = yueLyricsProblem(lyrics)
-    if (problem) throw createError({ statusCode: 400, statusMessage: problem })
-  }
   const duration = clampMusicDuration(body.duration)
-  if ((engine === 'yue' || engine === 'yue2') && duration > 150) throw createError({ statusCode: 400, statusMessage: 'YuE supports up to 150 seconds per section.' })
-  const steps = clampMusicSteps(
-    body.steps,
-    engine === 'ace-step-1.5' ? MUSIC_STEPS_DEFAULT_15 : MUSIC_STEPS_DEFAULT
-  )
-  const cfg = clampMusicCfg(
-    body.cfg,
-    engine === 'ace-step-1.5' ? MUSIC_CFG_DEFAULT_15 : MUSIC_CFG_DEFAULT
-  )
-  const lyricsStrength = clampLyricsStrength(body.lyricsStrength)
   const seed = parseMusicSeed(body.seed)
   const clipName = allocateOutputName(ownerKey, folderId, String(body.name || '').trim(), {
-    engine: engine === 'ace-step-1.5' ? 'ace15' : engine === 'yue' ? 'yue' : engine === 'yue2' ? 'yue2' : 'ace'
+    engine: 'yue2'
   })
 
-  const samplerName = String(body.sampler_name || 'euler')
-  const scheduler = String(body.scheduler || 'simple')
   const familyId = crypto.randomUUID()
   const studio = await addStudioJob({
     ownerKey,
@@ -105,15 +68,15 @@ export default defineEventHandler(async (event) => {
       aspect: 'audio',
       width: 0,
       height: 0,
-      steps,
+      steps: 0,
       turbo: false,
       seed,
-      cfg,
+      cfg: 0,
       fps: 0,
-      samplerName,
-      scheduler,
+      samplerName: 'euler',
+      scheduler: 'simple',
       duration,
-      sound: !instrumental,
+      sound: true,
       workflow: 'v1',
       useIdentityRefs: false,
       hideThumbnail: false,
@@ -122,10 +85,9 @@ export default defineEventHandler(async (event) => {
       extensions: [],
       queueAutoRun: false,
       lyrics,
-      instrumental,
-      lyricsStrength,
-      musicEngine: engine,
-      ...(engine === 'yue' ? { yueProfile: yueProfile as 1 | 3 } : {})
+      instrumental: false,
+      lyricsStrength: 0,
+      musicEngine: 'yue2'
     }
   })
   await kickStudioQueue()
@@ -142,11 +104,9 @@ export default defineEventHandler(async (event) => {
     queued: !liveJobId,
     seed,
     duration,
-    steps,
-    cfg,
     folderId,
     folderLocked,
-    engine,
+    engine: 'yue2' as const,
     name: clipName
   }
 })

server/api/health.get.ts

@@ -48,11 +48,21 @@ export default defineEventHandler(async () => {
       host: comfy.host || '',
       queueRunning: queue.running,
       queuePending: queue.pending,
-      asleep: Boolean(control?.asleep) && !comfyReachable,
+      asleep: Boolean(control?.asleep) && !comfyReachable && !control?.yue2?.busy && !control?.caption?.busy,
       processUp: Boolean(control?.process ?? control?.processUp),
       lastActivityAt: control?.lastActivityAt || null,
       stats: comfy.ok ? comfy.stats : undefined
     },
+    yue2: {
+      configured: Boolean(control?.yue2?.configured),
+      busy: Boolean(control?.yue2?.busy),
+      backend: control?.yue2?.backend || 'yue2'
+    },
+    caption: {
+      configured: Boolean(control?.caption?.configured),
+      busy: Boolean(control?.caption?.busy),
+      backend: control?.caption?.backend || 'llama.cpp'
+    },
     imageComfy: {
       configured: imageConfigured,
       ok: imageOk,

server/api/studio-2/caption.post.ts

@@ -0,0 +1,74 @@
+import { validateCaptionRequest } from '~/shared/studio2/caption.mjs'
+import { saveRecord, readRecord } from '../../utils/studio2/store'
+import { addStudioJob, kickStudioQueue, type StudioJobPayload } from '../../utils/studioQueue'
+import { getStill } from '../../utils/library'
+
+export default defineEventHandler(async (event) => {
+  const { owner } = assertLibraryOwner(event)
+  const raw = await readBody(event)
+  let request: ReturnType<typeof validateCaptionRequest>
+  try {
+    request = validateCaptionRequest(raw)
+  } catch (e: any) {
+    throw createError({ statusCode: e.statusCode || 400, statusMessage: e.message })
+  }
+  assertFolderAccess(event, request.folderId)
+  let stillId = request.stillId
+  if (!stillId && request.sourcePath) {
+    // sourcePath may be a still id when the client already resolved a library asset.
+    stillId = String(request.sourcePath)
+  }
+  const still = getStill(owner, stillId)
+  assertFolderAccess(event, still.folderId)
+  request.stillId = still.id
+  request.imageAId = still.id
+  request.sourcePath = still.id
+
+  const id = crypto.randomUUID()
+  const record = {
+    id,
+    owner,
+    request,
+    state: 'waiting',
+    queuedAt: Date.now(),
+    familyId: crypto.randomUUID(),
+    outputs: [],
+    resultText: '',
+    purgeResult: 'Not yet saved'
+  }
+  saveRecord(record)
+  const row = await addStudioJob({
+    ownerKey: owner,
+    kind: 'caption',
+    familyId: record.familyId,
+    payload: {
+      studio2Id: id,
+      prompt: request.compiledPrompt,
+      name: `Describe · ${request.captionStyle}`,
+      folderId: request.folderId,
+      stillId: still.id,
+      captionStyle: request.captionStyle,
+      aspect: 'auto',
+      width: 0,
+      height: 0,
+      steps: 1,
+      turbo: true,
+      seed: 0,
+      cfg: 0,
+      fps: 24,
+      samplerName: 'euler',
+      scheduler: 'simple',
+      duration: 0,
+      sound: false,
+      workflow: 'minimax',
+      useIdentityRefs: false,
+      hideThumbnail: false,
+      referenceStillIds: [],
+      extensions: [],
+      queueAutoRun: true
+    } as StudioJobPayload
+  })
+  saveRecord({ ...readRecord(owner, id), queueId: row.id })
+  kickStudioQueue()
+  return { id, captionStyle: request.captionStyle, stillId: still.id }
+})

server/api/studio-2/jobs.post.ts

@@ -14,8 +14,6 @@ export default defineEventHandler(async (event) => {
         throw createError({ statusCode: e.statusCode || 400, statusMessage: e.message });
     }
     const video = ['video', 'extend'].includes(request.mode);
-    if (video && request.engine !== 'minimax')
-        throw createError({ statusCode: 400, statusMessage: 'Use the existing xAIGen studio for LTX. Studio 2 video currently supports MiniMax.' });
     request.settings.loraStack = parsePostedLoraStack(request.settings.loraStack, video ? 'video' : 'image');
     request.settings.loraStack = filterLoraStackForUse(request.settings.loraStack, request.engine, useRuntimeConfig().public.studio === 'xaigen');
     if (!video)

server/api/studio-2/library.get.ts

@@ -5,5 +5,5 @@ export default defineEventHandler(event => {
  const { owner } = assertLibraryOwner(event);
  const library = publicLibrary(event);
  const catalog=studio2LoraCatalog(owner);
- return {...library,loraAliases:catalog.aliases,loraTriggers:catalog.triggers, stills:library.stills.filter(s=>existsSync(stillPath(owner,s.id))), clips:library.clips.filter(c=>existsSync(clipVideoPath(owner,c.id)))};
+ return {...library,loraAliases:catalog.aliases,loraTriggers:catalog.triggers, stills:library.stills.filter(s=>existsSync(stillPath(owner,s.id))), clips:library.clips.filter(c=>existsSync(clipVideoPath(owner,c.id))), tracks:library.tracks || []};
 });

server/api/studio-2/loras.get.ts

@@ -1,15 +1,19 @@
-import { readFileSync,writeFileSync,renameSync } from 'node:fs';
-import { join } from 'node:path';
-import { cachedLoras } from '~/shared/studio2/lora-cache.mjs';
-import { listStudioLoras } from '../../utils/loras';
-import { studio2Root } from '../../utils/studio2/store';
-import { studio2LoraCatalog } from '../../utils/studio2/lora-catalog';
+import { readFileSync,writeFileSync,renameSync } from 'node:fs'
+import { join } from 'node:path'
+import { cachedLoras } from '~/shared/studio2/lora-cache.mjs'
+import { listStudioLoras } from '../../utils/loras'
+import { studio2Root } from '../../utils/studio2/store'
+import { studio2LoraCatalog } from '../../utils/studio2/lora-catalog'
+
 export default defineEventHandler(async event => {
- const {owner}=assertLibraryOwner(event),path=join(studio2Root(),`loras-${owner}.cache`);
+  const { owner } = assertLibraryOwner(event)
+  const path = join(studio2Root(), `loras-${owner}.cache`)
+  // Same live Comfy scan as Settings (/api/loras?fresh=1). Always fresh so CACHE_MS cannot freeze adds.
   const listed = await cachedLoras({
-  read:()=>{try{return JSON.parse(readFileSync(path,'utf8'));}catch{return null;}},
-  write:value=>{writeFileSync(path+'.tmp',JSON.stringify(value));renameSync(path+'.tmp',path);},
-  live:()=>listStudioLoras({fresh:true}),fresh:getQuery(event).fresh==='1'
- });
- return {...listed,...studio2LoraCatalog(owner)};
-});
+    read: () => { try { return JSON.parse(readFileSync(path, 'utf8')) } catch { return null } },
+    write: value => { writeFileSync(path + '.tmp', JSON.stringify(value)); renameSync(path + '.tmp', path) },
+    live: () => listStudioLoras({ fresh: true }),
+    fresh: true
+  })
+  return { ...listed, ...studio2LoraCatalog(owner) }
+})

server/api/studio-queue/[id].patch.ts

@@ -50,22 +50,24 @@ export default defineEventHandler(async (event) => {
       payload.prompt = text(body.tags, 2000)
       row.prompt = payload.prompt
     }
-    if (body.yueProfile != null && row.kind === 'music' && payload.musicEngine === 'yue') {
-      if (body.yueProfile !== 1 && body.yueProfile !== 3) throw createError({ statusCode: 400, statusMessage: 'Choose YuEGP profile 1 or 3.' })
-      payload.yueProfile = body.yueProfile
-    }
     if (body.lyrics != null) payload.lyrics = text(body.lyrics)
     const instrumental = optionalBool(body.instrumental)
-    if (instrumental != null) payload.instrumental = instrumental
+    if (instrumental != null && row.kind === 'music') {
+      if (instrumental) throw createError({ statusCode: 400, statusMessage: 'YuE2 requires lyrics. Instrumental mode is not supported.' })
+      payload.instrumental = false
+    } else if (instrumental != null) {
+      payload.instrumental = instrumental
+    }
     if (body.promptPre != null) payload.promptPre = text(body.promptPre)
     if (body.promptMid != null) payload.promptMid = text(body.promptMid)
     if (body.promptPost != null) payload.promptPost = text(body.promptPost)
     if (body.negative != null) payload.negative = text(body.negative, 2000)
     const duration = optionalNumber(body.duration)
     if (duration != null) {
-      if (row.kind === 'music' && payload.musicEngine === 'yue') {
-        if (!Number.isInteger(duration) || duration < 30 || duration > 150) throw createError({ statusCode: 400, statusMessage: 'YuEGP duration must be 30–150 seconds.' })
+      if (row.kind === 'music') {
+        if (!Number.isInteger(duration) || duration < 15 || duration > 150) throw createError({ statusCode: 400, statusMessage: 'YuE2 target length must be 15–150 seconds.' })
         payload.duration = duration
+        payload.musicEngine = 'yue2'
       } else payload.duration = Math.min(120, Math.max(0.5, duration))
     }
     const steps = optionalNumber(body.steps)

server/assets/studio2_qwen21_t2i.json

@@ -0,0 +1,75 @@
+{
+  "4": {
+    "inputs": {
+      "unet_name": "qwen-image-2.1-Q8_0.gguf"
+    },
+    "class_type": "UnetLoaderGGUF",
+    "_meta": { "title": "Load Qwen 2.1 GGUF" }
+  },
+  "3": {
+    "inputs": {
+      "shift": 3.1,
+      "model": ["4", 0]
+    },
+    "class_type": "ModelSamplingAuraFlow",
+    "_meta": { "title": "Qwen 2.1 shift" }
+  },
+  "5": {
+    "inputs": {
+      "clip_name": "qwen3vl_8b_int8_convrot.safetensors",
+      "type": "qwen_image",
+      "device": "default"
+    },
+    "class_type": "CLIPLoader",
+    "_meta": { "title": "Load Qwen 2.1 CLIP" }
+  },
+  "6": {
+    "inputs": {
+      "vae_name": "qwen_image_2.1_vae_bf16.safetensors"
+    },
+    "class_type": "VAELoader",
+    "_meta": { "title": "Load Qwen 2.1 VAE" }
+  },
+  "9": {
+    "inputs": {
+      "clip": ["5", 0],
+      "prompt": "",
+      "negative_prompt": "",
+      "resolution": 1024
+    },
+    "class_type": "TextEncodeQwenImage21",
+    "_meta": { "title": "Text Encode Qwen Image 2.1" }
+  },
+  "15": {
+    "inputs": {
+      "seed": 1,
+      "steps": 25,
+      "cfg": 1,
+      "sampler_name": "euler",
+      "scheduler": "simple",
+      "denoise": 1,
+      "model": ["3", 0],
+      "positive": ["9", 0],
+      "negative": ["9", 1],
+      "latent_image": ["9", 2]
+    },
+    "class_type": "KSampler",
+    "_meta": { "title": "KSampler" }
+  },
+  "20": {
+    "inputs": {
+      "samples": ["15", 0],
+      "vae": ["6", 0]
+    },
+    "class_type": "VAEDecode",
+    "_meta": { "title": "VAE Decode" }
+  },
+  "21": {
+    "inputs": {
+      "filename_prefix": "studio2-qwen21",
+      "images": ["20", 0]
+    },
+    "class_type": "SaveImage",
+    "_meta": { "title": "Save Image" }
+  }
+}

server/plugins/00-resume-caption.ts

@@ -0,0 +1,3 @@
+import { resumeCaptionJobs } from '~/server/utils/caption'
+
+export default defineNitroPlugin(() => { resumeCaptionJobs() })

server/plugins/00-resume-yuegp.ts

@@ -1,4 +1,2 @@
-import { resumeYueGpJobs } from '~/server/utils/yueGp'
-
-// Restore live music IDs before queue repair examines their durable studio rows.
-export default defineNitroPlugin(() => { resumeYueGpJobs() })
+// YuEGP resume removed — music is YuE2-only. scripts/yuegp-* remain on disk unused.
+export default defineNitroPlugin(() => {})

server/utils/caption.ts

@@ -0,0 +1,268 @@
+import { createReadStream, existsSync, mkdirSync, readdirSync, readFileSync, writeFileSync, renameSync, unlinkSync } from 'node:fs'
+import { join } from 'node:path'
+import { createJob, emitJob, getJob, restoreJob, type Job } from './jobs'
+import { getStill, stillPath } from './library'
+import { sharedGpuHeaders } from './sharedGpu'
+import { CAPTION_STYLES } from '~/shared/studio2/caption.mjs'
+
+function settings() {
+  const config = useRuntimeConfig()
+  return {
+    url: String(config.comfyControlUrl || process.env.COMFY_CONTROL_URL || '').replace(/\/$/, ''),
+    token: String(config.comfyControlToken || process.env.COMFY_CONTROL_TOKEN || '')
+  }
+}
+
+export function captionConfigured() { return Boolean(settings().url) }
+
+function pendingRoot() {
+  return join(String(useRuntimeConfig().libraryDir || process.env.LIBRARY_DIR || '/data/library'), 'caption-pending')
+}
+
+export function captionPendingAlive(jobId: string) {
+  if (!jobId) return false
+  return existsSync(join(pendingRoot(), `${jobId}.json`))
+}
+
+async function request(path: string, method = 'GET', body?: unknown) {
+  const { url, token } = settings()
+  if (!url) throw new Error('Caption host is not configured. Set COMFY_CONTROL_URL.')
+  const response = await fetch(`${url}/caption/${path}`, {
+    method,
+    headers: {
+      ...(token ? { Authorization: `Bearer ${token}` } : {}),
+      ...(method === 'GET' ? {} : sharedGpuHeaders()),
+      ...(method === 'POST' ? { 'Content-Type': 'application/json' } : {})
+    },
+    body: method === 'POST' ? JSON.stringify(body) : body as BodyInit | undefined,
+    ...(method === 'PUT' ? { duplex: 'half' as const } : {}),
+    signal: AbortSignal.timeout(method === 'PUT' ? 120_000 : 60_000)
+  } as RequestInit)
+  if (!response.ok) {
+    const detail = await response.json().catch(() => ({})) as { message?: string; error?: string }
+    throw Object.assign(new Error(detail.message || detail.error || `Caption host returned ${response.status}`), { statusCode: response.status })
+  }
+  return response
+}
+
+function persist(record: Record<string, unknown>) {
+  mkdirSync(pendingRoot(), { recursive: true })
+  const path = join(pendingRoot(), `${record.id}.json`)
+  writeFileSync(path + '.tmp', JSON.stringify(record))
+  renameSync(path + '.tmp', path)
+}
+
+function clearPending(id: string) {
+  const path = join(pendingRoot(), `${id}.json`)
+  if (existsSync(path)) unlinkSync(path)
+}
+
+function writeSidecar(owner: string, stillId: string, text: string) {
+  try {
+    const image = stillPath(owner, stillId)
+    if (!existsSync(image)) return
+    writeFileSync(`${image}.txt`, text, 'utf8')
+  } catch { /* optional sidecar */ }
+}
+
+export async function cancelCaptionJob(job: Job) {
+  await request(`jobs/${job.id}/cancel`, 'POST', {})
+  job.status = 'cancelled'
+  emitJob(job, { type: 'error', error: 'Cancelled', message: 'Cancelled' })
+}
+
+async function settle(job: Job) {
+  const { onLiveVideoSettled } = await import('./studioQueue')
+  await onLiveVideoSettled(job)
+  clearPending(job.id)
+}
+
+async function watch(job: Job, record: {
+  id: string
+  owner: string
+  stillId: string
+  folderId: string
+  captionStyle: string
+  studio2Id?: string
+  liveId: string
+}) {
+  let failures = 0
+  while (job.status !== 'cancelled') {
+    try {
+      const state = await (await request(`jobs/${job.id}`)).json() as {
+        status: string
+        message?: string
+        error?: string
+        progress?: number
+        text?: string
+      }
+      if (job.status === 'cancelled') { await settle(job); return }
+      if (state.status === 'error' || state.status === 'cancelled') {
+        job.status = state.status === 'cancelled' ? 'cancelled' : 'error'
+        job.error = state.error || state.message || 'Caption failed'
+        emitJob(job, { type: 'error', error: job.error, message: job.error })
+        if (record.studio2Id) {
+          const { readRecord, saveRecord } = await import('./studio2/store')
+          try {
+            const r = readRecord(record.owner, record.studio2Id)
+            r.state = job.status === 'cancelled' ? 'cancelled' : 'failed'
+            r.error = job.error
+            r.finishedAt = Date.now()
+            saveRecord(r)
+          } catch { /* record may already be gone */ }
+        }
+        await settle(job)
+        return
+      }
+      if (state.status === 'complete') {
+        const text = String(state.text || '').trim()
+        job.message = 'Caption ready'
+        job.progress = 100
+        job.status = 'complete'
+        ;(job as Job & { resultText?: string }).resultText = text
+        if (record.studio2Id) {
+          const { readRecord, saveRecord } = await import('./studio2/store')
+          const r = readRecord(record.owner, record.studio2Id)
+          r.state = 'complete'
+          r.resultText = text
+          r.finishedAt = Date.now()
+          r.savedAt = Date.now()
+          saveRecord(r)
+        }
+        writeSidecar(record.owner, record.stillId, text)
+        emitJob(job, { type: 'complete', message: 'Caption ready', progress: 100 })
+        await settle(job)
+        return
+      }
+      failures = 0
+      job.status = 'running'
+      emitJob(job, {
+        type: 'progress',
+        message: state.message || 'Captioning',
+        progress: Math.min(95, Math.max(1, Number(state.progress || 10)))
+      })
+    } catch (error) {
+      failures++
+      emitJob(job, { type: 'status', message: `Caption host check failed; retrying: ${error instanceof Error ? error.message : String(error)}` })
+      if (failures >= 10) {
+        job.status = 'error'
+        job.error = 'Caption host unreachable.'
+        emitJob(job, { type: 'error', error: job.error, message: job.error })
+        await settle(job)
+        return
+      }
+    }
+    await new Promise(resolve => setTimeout(resolve, 1500))
+  }
+  await settle(job)
+}
+
+export async function startCaptionJob(params: {
+  ownerKey: string
+  folderId: string
+  stillId: string
+  captionStyle: string
+  studio2Id?: string
+  name?: string
+}) {
+  if (!CAPTION_STYLES.includes(params.captionStyle)) throw new Error('Unknown caption style.')
+  getStill(params.ownerKey, params.stillId)
+  const job = createJob('caption')
+  job.caption = true
+  job.library = {
+    ownerKey: params.ownerKey,
+    folderId: params.folderId,
+    hideThumbnail: false,
+    name: params.name || `Describe · ${params.captionStyle}`,
+    prompt: params.captionStyle,
+    aspect: 'image',
+    width: 0,
+    height: 0,
+    steps: 1,
+    turbo: true,
+    seed: 0,
+    stillId: params.stillId,
+    engine: 'caption'
+  }
+  const record = {
+    id: job.id,
+    liveId: job.id,
+    owner: params.ownerKey,
+    stillId: params.stillId,
+    folderId: params.folderId,
+    captionStyle: params.captionStyle,
+    studio2Id: params.studio2Id,
+    createdAt: Date.now(),
+    status: 'running'
+  }
+  persist(record)
+  void (async () => {
+    try {
+      if (job.status === 'cancelled') { await settle(job); return }
+      await request(`jobs/${job.id}/input`, 'PUT', createReadStream(stillPath(params.ownerKey, params.stillId)))
+      if (job.status === 'cancelled') { await settle(job); return }
+      try {
+        await request('jobs', 'POST', { id: job.id, style: params.captionStyle })
+      } catch {
+        /* Confirm by ID; never caption twice on an HTTP timeout. */
+      }
+      await watch(job, record)
+    } catch (error) {
+      const statusCode = (error as { statusCode?: number })?.statusCode
+      if (statusCode && statusCode >= 400 && statusCode < 500) {
+        job.status = 'error'
+        job.error = error instanceof Error ? error.message : String(error)
+        emitJob(job, { type: 'error', error: job.error, message: job.error })
+        if (params.studio2Id) {
+          try {
+            const { readRecord, saveRecord } = await import('./studio2/store')
+            const r = readRecord(params.ownerKey, params.studio2Id)
+            r.state = 'failed'
+            r.error = job.error
+            r.finishedAt = Date.now()
+            saveRecord(r)
+          } catch { /* ignore */ }
+        }
+        await settle(job)
+        return
+      }
+      emitJob(job, { type: 'status', message: `Checking caption submission: ${error instanceof Error ? error.message : String(error)}` })
+      await watch(job, record)
+    }
+  })()
+  return job
+}
+
+export function resumeCaptionJobs() {
+  if (!existsSync(pendingRoot())) return
+  for (const file of readdirSync(pendingRoot()).filter(file => /^[a-zA-Z0-9-]+\.json$/.test(file))) {
+    try {
+      const record = JSON.parse(readFileSync(join(pendingRoot(), file), 'utf8'))
+      if (getJob(record.id)) continue
+      const job = restoreJob({
+        id: record.liveId || record.id,
+        clientId: record.clientId || crypto.randomUUID(),
+        promptId: '',
+        startedAt: record.createdAt || Date.now(),
+        library: {
+          ownerKey: record.owner,
+          folderId: record.folderId,
+          hideThumbnail: false,
+          prompt: record.captionStyle || 'descriptive',
+          aspect: 'image',
+          width: 0,
+          height: 0,
+          steps: 1,
+          turbo: true,
+          seed: 0,
+          stillId: record.stillId,
+          engine: 'caption'
+        }
+      })
+      job.caption = true
+      job.kind = 'caption'
+      job.message = 'Reconnecting to caption host'
+      void watch(job, record)
+    } catch { /* Preserve invalid records for diagnosis. */ }
+  }
+}

server/utils/generationPresets.ts

@@ -166,17 +166,12 @@ function sanitizeImageV2Settings(raw: unknown): ImageV2PresetSettings {
 
 function sanitizeMusicSettings(raw: unknown): MusicPresetSettings {
   const rec = raw && typeof raw === 'object' ? raw as Record<string, unknown> : {}
-  const instrumental = rec.instrumental === true
   return {
     tags: sanitizeMusicTags(rec.tags),
-    lyrics: sanitizeMusicLyrics(rec.lyrics, instrumental),
-    instrumental,
+    lyrics: sanitizeMusicLyrics(rec.lyrics, false),
     duration: clampMusicDuration(rec.duration),
-    steps: clampMusicSteps(rec.steps),
     seed: typeof rec.seed === 'number' && Number.isFinite(rec.seed) ? Math.max(0, Math.floor(rec.seed)) : undefined,
-    cfg: rec.cfg == null ? undefined : clampMusicCfg(rec.cfg),
-    engine: parseMusicEngine(rec.engine),
-    ...(parseMusicEngine(rec.engine) === 'yue' ? { yueProfile: rec.yueProfile === 3 ? 3 as const : 1 as const } : {})
+    engine: 'yue2'
   }
 }
 

server/utils/jobs.ts

@@ -36,9 +36,11 @@ export interface Job {
   upscale?: boolean
   yueGp?: boolean
   yue2?: boolean
+  caption?: boolean
+  resultText?: string
   musicActivity?: { checkedAt: number; running: boolean }
   id: string
-  kind?: 'video' | 'edit' | 'music'
+  kind?: 'video' | 'edit' | 'music' | 'caption'
   promptId?: string
   clientId: string
   status: JobStatus

server/utils/loras.ts

@@ -44,6 +44,7 @@ type LoraCache = {
 let cache: LoraCache | null = null
 let inflight: Promise<LoraCache> | null = null
 let inflightFresh = false
+let loadSeq = 0
 
 function uniqueNames(values: unknown[]) {
   const seen = new Set<string>()
@@ -80,14 +81,44 @@ function namesFromObjectInfoNode(info: unknown) {
   return uniqueNames(names)
 }
 
-async function fetchJson(path: string, via: 'video' | 'image') {
+async function fetchJson(path: string, via: 'video' | 'image', timeoutMs = 8000) {
   const res = via === 'image'
-    ? await imageComfyFetch(path, { signal: AbortSignal.timeout(8000) })
-    : await comfyFetch(path, { signal: AbortSignal.timeout(8000) })
+    ? await imageComfyFetch(path, { signal: AbortSignal.timeout(timeoutMs) })
+    : await comfyFetch(path, { signal: AbortSignal.timeout(timeoutMs) })
   if (!res.ok) return null
   return res.json().catch(() => null)
 }
 
+function collectObjectInfoLoras(all: Record<string, unknown>) {
+  const image: string[] = []
+  const video: string[] = []
+  let any = false
+  for (const node of OBJECT_INFO_NODES) {
+    const info = all[node]
+    if (!info || typeof info !== 'object') continue
+    const names = namesFromObjectInfoNode(info)
+    if (!names.length) continue
+    any = true
+    if (node === LORA_LOADER) image.push(...names)
+    else video.push(...names)
+    if (node !== LORA_LOADER && node !== LORA_MODEL_ONLY) {
+      image.push(...names)
+      video.push(...names)
+    }
+  }
+  if (any) return { image: uniqueNames(image), video: uniqueNames(video) }
+
+  const imageAll: string[] = []
+  const videoAll: string[] = []
+  for (const [classType, info] of Object.entries(all)) {
+    if (!/lora/i.test(classType)) continue
+    const names = namesFromObjectInfoNode(info)
+    if (classType === LORA_LOADER) imageAll.push(...names)
+    else videoAll.push(...names)
+  }
+  return { image: uniqueNames(imageAll), video: uniqueNames(videoAll) }
+}
+
 async function fetchModelsLoras(via: 'video' | 'image') {
   const payload = await fetchJson('/models/loras', via)
   if (Array.isArray(payload)) return uniqueNames(payload)
@@ -97,7 +128,14 @@ async function fetchModelsLoras(via: 'video' | 'image') {
   return [] as string[]
 }
 
-async function fetchObjectInfoLoras(via: 'video' | 'image') {
+async function fetchObjectInfoLoras(via: 'video' | 'image', fresh = false) {
+  // Full /object_info is Comfy's model-list refresh (UI Refresh) and lists LoraLoader / model-only / rgthree / WanVideoLoraSelect.
+  if (fresh) {
+    const all = await fetchJson('/object_info', via, 45_000)
+    if (all && typeof all === 'object') return collectObjectInfoLoras(all as Record<string, unknown>)
+    return { image: [] as string[], video: [] as string[] }
+  }
+
   const image: string[] = []
   const video: string[] = []
   let any = false
@@ -120,22 +158,14 @@ async function fetchObjectInfoLoras(via: 'video' | 'image') {
     return { image: uniqueNames(image), video: uniqueNames(video) }
   }
 
-  const all = await fetchJson('/object_info', via)
+  const all = await fetchJson('/object_info', via, 45_000)
   if (!all || typeof all !== 'object') return { image: [] as string[], video: [] as string[] }
-  const imageAll: string[] = []
-  const videoAll: string[] = []
-  for (const [classType, info] of Object.entries(all as Record<string, unknown>)) {
-    if (!/lora/i.test(classType)) continue
-    const names = namesFromObjectInfoNode(info)
-    if (classType === LORA_LOADER) imageAll.push(...names)
-    else videoAll.push(...names)
-  }
-  return { image: uniqueNames(imageAll), video: uniqueNames(videoAll) }
+  return collectObjectInfoLoras(all as Record<string, unknown>)
 }
 
-async function discoverFromHost(via: 'video' | 'image') {
+async function discoverFromHost(via: 'video' | 'image', fresh = false) {
+  const fromInfo = await fetchObjectInfoLoras(via, fresh).catch(() => ({ image: [] as string[], video: [] as string[] }))
   const models = await fetchModelsLoras(via).catch(() => [] as string[])
-  const fromInfo = await fetchObjectInfoLoras(via).catch(() => ({ image: [] as string[], video: [] as string[] }))
   const image = uniqueNames([...fromInfo.image, ...models])
   const video = uniqueNames([...fromInfo.video, ...models])
   if (!image.length && video.length) return { image: video, video }
@@ -147,17 +177,19 @@ async function loadLoraCache(options: { fresh?: boolean } = {}): Promise<LoraCac
   const fresh = options.fresh === true
   const now = Date.now()
   if (!fresh && cache && now - cache.at < CACHE_MS) return cache
+  // Fresh must not reuse a non-fresh inflight (stale list / frozen CACHE_MS window).
   if (inflight && (!fresh || inflightFresh)) return inflight
+  const seq = ++loadSeq
   inflightFresh = fresh
   inflight = (async () => {
     const videoHost = comfyConfigured() ? getComfyHost() : ''
     const imageHost = getBeastImageHost()
     const same = Boolean(videoHost && imageHost && sameImageHost(videoHost, imageHost))
     const video = videoHost
-      ? await discoverFromHost('video').catch(() => ({ image: [] as string[], video: [] as string[] }))
+      ? await discoverFromHost('video', fresh).catch(() => ({ image: [] as string[], video: [] as string[] }))
       : { image: [] as string[], video: [] as string[] }
     const image = imageHost && !same
-      ? await discoverFromHost('image').catch(() => ({ image: [] as string[], video: [] as string[] }))
+      ? await discoverFromHost('image', fresh).catch(() => ({ image: [] as string[], video: [] as string[] }))
       : video
     const next: LoraCache = {
       at: Date.now(),
@@ -166,15 +198,19 @@ async function loadLoraCache(options: { fresh?: boolean } = {}): Promise<LoraCac
     }
     // Comfy asleep / unreachable returns empty. Keep the last good list so the picker
     // does not vanish — but do NOT refresh `at`, or a blip freezes new folder drops out.
-    // Never merge a shorter fresh list back into the old one: deletes must disappear.
+    // Never merge a shorter live list into the old one: deletes must disappear; adds must appear.
     if (!next.image.length && !next.video.length && cache && (cache.image.length || cache.video.length)) {
       return cache
     }
+    // Superseded non-fresh must not overwrite a newer fresh result.
+    if (seq !== loadSeq) return cache || next
     cache = next
     return next
   })().finally(() => {
+    if (seq === loadSeq) {
       inflight = null
       inflightFresh = false
+    }
   })
   return inflight
 }

server/utils/musicChain.ts

@@ -1,13 +1,4 @@
-import { startYueGpJob } from './yueGp'
 import { startYue2Job } from './yue2'
-import { createJob, emitJob, type Job } from '~/server/utils/jobs'
-import { extractAudio, fetchHistory, fetchHistoryAll, findHistoryAudio, purgeComfyArtifacts, queuePrompt } from '~/server/utils/comfy'
-import { comfyWsUrl, comfyFetch } from '~/server/utils/comfy'
-import { ensureComfyReady } from '~/server/utils/comfyLifecycle'
-import { downloadComfyAudio, saveTrack } from '~/server/utils/library'
-import { buildMusicWorkflow, musicFilenamePrefix, assertMusicEngineNodes } from '~/server/utils/musicWorkflow'
-import { NODE_LABELS } from '~/server/utils/workflow'
-import { musicExtension } from '~/utils/music'
 import type { MusicEngine } from '~/utils/music'
 
 export type MusicJobParams = {
@@ -23,413 +14,22 @@ export type MusicJobParams = {
   lyricsStrength: number
   instrumental: boolean
   folderLocked?: boolean
-  yueProfile?: 1 | 3
   engine?: MusicEngine
   samplerName?: string
   scheduler?: string
 }
 
-function sleep(ms: number) {
-  return new Promise(resolve => setTimeout(resolve, ms))
-}
-
-function classifyError(message: string) {
-  const lower = message.toLowerCase()
-  if (lower.includes('out of memory') || lower.includes('exceed allowed memory') || (lower.includes('cuda') && lower.includes('alloc')) || lower.includes('vram')) {
-    return 'ComfyUI ran out of VRAM.'
-  }
-  if (lower.includes('econnrefused') || lower.includes('unreachable') || lower.includes('fetch failed')) {
-    return 'ComfyUI host connection dropped. Confirm the desktop instance is running.'
-  }
-  if (lower.includes('timeout')) {
-    return 'Network timeout talking to ComfyUI. The job may still be running on the desktop.'
-  }
-  return message
-}
-
-async function settleMusicStudio(job: Job) {
-  try {
-    const { onLiveVideoSettled } = await import('~/server/utils/studioQueue')
-    await onLiveVideoSettled(job)
-  } catch {
-    /* queue kick is best-effort */
-  }
-}
-
-function nodeLabel(node: string) {
-  if (!node) return ''
-  return NODE_LABELS[node] || ''
-}
-
-async function waitForSocket(job: Job, ms = 4000) {
-  const started = Date.now()
-  while (Date.now() - started < ms) {
-    if (job.socketReady) return
-    await sleep(100)
-  }
-}
-
-function watchMusicJob(job: Job): Promise<void> {
-  job.socketReady = false
-  job.saving = false
-  const ws = new WebSocket(comfyWsUrl(job.clientId))
-  let settled = false
-  let finishing = false
-  let pollTimer: ReturnType<typeof setInterval> | null = null
-  let timeout: ReturnType<typeof setTimeout> | null = null
-
-  return new Promise((resolve) => {
-    let localProgress = 0
-    const startedAt = Date.now()
-    let lastActivity = Date.now()
-    const IDLE_MS = 30 * 60 * 1000
-    const ABSOLUTE_MS = 3 * 60 * 60 * 1000
-
-    const markActivity = () => { lastActivity = Date.now() }
-
-    const emitLocal = (event: Parameters<typeof emitJob>[1], samplePct?: number) => {
-      markActivity()
-      if (typeof event.progress === 'number') localProgress = event.progress
-      if (event.message && typeof samplePct === 'number') {
-        event.message = `${event.message} (${samplePct}%)`
-      }
-      emitJob(job, event)
-    }
-
-    const cleanup = () => {
-      if (timeout) clearTimeout(timeout)
-      if (pollTimer) clearInterval(pollTimer)
-      timeout = null
-      pollTimer = null
-      try { ws.close() } catch { /* ignore */ }
-    }
-
-    const armIdle = () => {
-      if (timeout) clearTimeout(timeout)
-      if (settled) return
-      const absLeft = ABSOLUTE_MS - (Date.now() - startedAt)
-      if (absLeft <= 0) {
-        void fail('Timed out waiting for ComfyUI (3 hours).')
-        return
-      }
-      const idleLeft = IDLE_MS - (Date.now() - lastActivity)
-      timeout = setTimeout(() => {
-        if (settled || finishing) return
-        if (Date.now() - lastActivity >= IDLE_MS) {
-          void fail('Timed out waiting for ComfyUI progress (30 minutes with no updates).')
-          return
-        }
-        armIdle()
-      }, Math.max(1000, Math.min(idleLeft, absLeft)))
-    }
-
-    const fail = async (error: string) => {
-      if (settled || finishing) return
-      settled = true
-      cleanup()
-      job.status = job.status === 'cancelled' ? 'cancelled' : 'error'
-      job.error = classifyError(error)
-      emitJob(job, { type: 'error', error: job.error, message: job.error })
-      await settleMusicStudio(job)
-      resolve()
-    }
-
-    const adoptPromptId = (id: unknown) => {
-      const promptId = String(id || '').trim()
-      if (promptId && !job.promptId) job.promptId = promptId
-    }
-
-    const loadJobAudio = async () => {
-      if (job.promptId) {
-        const scoped = await fetchHistory(job.promptId)
-        const fromPrompt = extractAudio(scoped, job.promptId)
-        if (fromPrompt) return fromPrompt
-      }
-      const found = findHistoryAudio(await fetchHistoryAll(), {
-        promptId: job.promptId,
-        tags: job.library?.tags,
-        since: job.startedAt - 60_000
-      })
-      if (found) {
-        job.promptId = found.promptId
-        return found.audio
-      }
-      return null
-    }
-
-    const succeed = async () => {
-      if (settled || finishing) return false
-      finishing = true
-      job.saving = true
-      try {
-        const audio = await loadJobAudio()
-        if (!audio) {
-          finishing = false
-          job.saving = false
-          return false
-        }
-        if (settled) return false
-        settled = true
-        cleanup()
-        try {
-          job.audio = audio
-          emitLocal({ type: 'status', message: 'Saving to library...', progress: 98 })
-          if (job.library) {
-            const buffer = await downloadComfyAudio(audio)
-            const ext = musicExtension(audio.filename)
-            const track = await saveTrack({
-              ownerKey: job.library.ownerKey,
-              folderId: job.library.folderId,
-              name: job.library.name,
-              tags: job.library.tags || job.library.prompt,
-              lyrics: job.library.lyrics || '',
-              duration: job.library.duration || 0,
-              steps: job.library.steps,
-              seed: job.library.seed,
-              cfg: job.library.cfg || 0,
-              instrumental: job.library.instrumental === true,
-              audio: buffer,
-              ext,
-              engine: job.library.engine,
-              samplerName: job.library.samplerName,
-              scheduler: job.library.scheduler,
-              lyricsStrength: job.library.lyricsStrength,
-              comfyFilename: audio.filename
-            })
-            job.trackId = track.id
-            job.library.audioExt = ext
-            await purgeComfyArtifacts({
-              audio,
-              promptId: job.promptId
-            })
-          }
-          job.status = 'complete'
-          emitJob(job, {
-            type: 'complete',
-            message: job.library?.folderLocked
-              ? 'Saved to the locked folder. Unlock it to listen.'
-              : 'Track ready',
-            progress: 100,
-            filename: job.library?.folderLocked ? undefined : audio.filename,
-            subfolder: job.library?.folderLocked ? undefined : audio.subfolder,
-            mediaType: job.library?.folderLocked ? undefined : audio.type,
-            trackId: job.trackId,
-            folderLocked: job.library?.folderLocked
-          })
-        } catch (saveError) {
-          const message = saveError instanceof Error ? saveError.message : String(saveError)
-          job.status = 'error'
-          job.error = `Audio generated but library save failed: ${message}`
-          emitJob(job, { type: 'error', error: job.error, message: job.error })
-        }
-        await settleMusicStudio(job)
-        job.saving = false
-        resolve()
-        return true
-      } catch (error) {
-        finishing = false
-        job.saving = false
-        throw error
-      }
-    }
-
-    const pollHistory = async () => {
-      if (settled || finishing) return
-      if (job.promptId) {
-        try {
-          const response = await comfyFetch('/queue', { signal: AbortSignal.timeout(2500) })
-          if (response.ok) {
-            const queue = await response.json() as { queue_running?: unknown[][] }
-            job.musicActivity = {
-              checkedAt: Date.now(),
-              running: Boolean(queue.queue_running?.some(row => row[1] === job.promptId))
-            }
-          }
-        } catch { /* Keep the last confirmation timestamp so the UI can show stale checks. */ }
-      }
-      if (settled || finishing) return
-      try {
-        const history = await fetchHistoryAll()
-        const found = findHistoryAudio(history, {
-          promptId: job.promptId,
-          tags: job.library?.tags,
-          since: job.startedAt - 60_000
-        })
-        if (found) {
-          job.promptId = found.promptId
-          await succeed()
-          return
-        }
-        const entry = (job.promptId ? history?.[job.promptId] : undefined) as {
-          status?: {
-            status_str?: string
-            completed?: boolean
-            messages?: Array<[string, Record<string, unknown>]>
-          }
-        } | undefined
-        const status = entry?.status?.status_str
-        if (status === 'error') {
-          const err = entry?.status?.messages?.find(([type]) => type === 'execution_error')?.[1]
-          const detail = String(err?.exception_message || err?.message || '').trim()
-          await fail(detail || 'ComfyUI reported an execution error')
-        } else if (status === 'interrupted') await fail('Job interrupted.')
-      } catch {
-        // history can miss while the graph is still running
-      }
-    }
-
-    armIdle()
-    pollTimer = setInterval(() => { void pollHistory() }, 4000)
-
-    ws.addEventListener('open', () => {
-      job.socketReady = true
-      emitLocal({ type: 'status', message: 'Connected to ComfyUI', progress: Math.max(localProgress, 4) })
-    })
-    ws.addEventListener('error', () => {
-      job.socketReady = false
-    })
-    ws.addEventListener('message', (event) => {
-      let payload: { type?: string; data?: Record<string, any> }
-      try {
-        payload = JSON.parse(String(event.data))
-      } catch {
-        return
-      }
-      const type = String(payload.type || '')
-      const data = payload.data || {}
-      adoptPromptId(data.prompt_id)
-      if (type === 'status') markActivity()
-      if (type === 'progress') {
-        const value = Number(data.value || 0)
-        const max = Number(data.max || 0)
-        const pct = max > 0 ? Math.min(95, Math.round((value / max) * 90) + 8) : Math.max(localProgress, 12)
-        const node = String(data.node || '')
-        emitLocal({
-          type: 'progress',
-          message: nodeLabel(node) || 'Sampling',
-          progress: pct,
-          step: value,
-          maxStep: max,
-          node
-        }, max > 0 ? Math.round((value / max) * 100) : undefined)
-      }
-      if (type === 'executing') {
-        const node = data.node == null ? '' : String(data.node)
-        if (!node) void succeed()
-        else if (node) {
-          emitLocal({
-            type: 'status',
-            message: nodeLabel(node) || 'Running node',
-            progress: Math.max(localProgress, 10),
-            node
-          })
-        }
-      }
-      if (type === 'executed') {
-        void succeed()
-      }
-      if (type === 'execution_error') {
-        const message = String(data.exception_message || data.message || 'ComfyUI node execution failed')
-        void fail(message)
-      }
-      if (type === 'execution_interrupted') {
-        job.status = 'cancelled'
-        void fail('Job interrupted.')
-      }
-    })
-  })
-}
-
+/** Music generation is YuE2-only. ACE Comfy and YuEGP paths are removed. */
 export async function startMusicJob(params: MusicJobParams) {
-  if (params.engine === 'yue') return startYueGpJob(params)
-  if (params.engine === 'yue2') return startYue2Job(params)
-  const job = createJob('music')
-  job.library = {
-    ownerKey: params.ownerKey,
-    folderId: params.folderId,
-    hideThumbnail: false,
-    folderLocked: params.folderLocked === true,
-    name: params.name,
-    prompt: params.tags,
-    tags: params.tags,
-    lyrics: params.lyrics,
-    instrumental: params.instrumental,
-    aspect: 'audio',
-    width: 0,
-    height: 0,
-    steps: params.steps,
-    turbo: false,
-    seed: params.seed,
-    cfg: params.cfg,
-    duration: params.duration,
-    sound: true,
-    engine: params.engine || 'ace-step',
-    samplerName: params.samplerName || 'euler',
-    scheduler: params.scheduler || 'simple',
-    lyricsStrength: params.lyricsStrength
+  if (params.instrumental) {
+    throw new Error('YuE2 requires lyrics. Instrumental mode is not supported.')
   }
-  job.maxStep = params.steps
-  job.status = 'queued'
-  emitJob(job, { type: 'status', message: 'Checking ComfyUI...', progress: 1 })
-
-  void (async () => {
-    try {
-      const done = watchMusicJob(job)
-      await ensureComfyReady((status) => {
-        emitJob(job, {
-          type: status.state === 'busy' ? 'busy' : 'status',
-          message: status.message,
-          progress: status.state === 'online' ? 3 : 1,
-          busy: status.state === 'busy',
-          queueRunning: status.queueRunning,
-          queuePending: status.queuePending
-        })
-      })
-      job.status = 'running'
-      const engineLabel = params.engine === 'ace-step-1.5' ? 'ACE-Step 1.5' : 'ACE-Step'
-      emitJob(job, { type: 'status', message: `Queueing ${engineLabel}...`, progress: 6 })
-      await waitForSocket(job, 4000)
-      await assertMusicEngineNodes(params.engine)
-      const graph = buildMusicWorkflow({
-        tags: params.tags,
-        lyrics: params.instrumental ? '' : params.lyrics,
-        duration: params.duration,
-        steps: params.steps,
-        seed: params.seed,
-        cfg: params.cfg,
-        lyricsStrength: params.instrumental ? 0 : params.lyricsStrength,
-        samplerName: params.samplerName,
-        scheduler: params.scheduler,
-        filenamePrefix: musicFilenamePrefix(),
-        engine: params.engine,
-        vocalMultiplier: params.instrumental ? 0.2 : 1
-      })
-      try {
-        const queued = await queuePrompt(graph, job.clientId, { timeoutMs: 25_000 })
-        if (queued?.prompt_id) job.promptId = queued.prompt_id
-        const runningMessage = 'Job queued on ComfyUI'
-        emitJob(job, { type: 'status', message: runningMessage, progress: 8 })
-      } catch (queueError: unknown) {
-        const statusCode = Number((queueError as { statusCode?: number; status?: number })?.statusCode
-          || (queueError as { status?: number })?.status || 0)
-        if (statusCode >= 400) throw queueError
-        emitJob(job, {
-          type: 'status',
-          message: 'Waiting for ComfyUI to pick up the job...',
-          progress: 8
+  if (!String(params.lyrics || '').trim()) {
+    throw new Error('YuE2 needs lyrics.')
+  }
+  return startYue2Job({
+    ...params,
+    engine: 'yue2',
+    instrumental: false
   })
 }
-      await done
-    } catch (error) {
-      const message = error instanceof Error ? error.message : String(error)
-      if (job.status !== 'cancelled' && job.status !== 'complete') {
-        job.status = 'error'
-        job.error = classifyError(message)
-        emitJob(job, { type: 'error', error: job.error, message: job.error })
-        await settleMusicStudio(job)
-      }
-    }
-  })()
-
-  return job
-}

server/utils/studio2/runner.ts

@@ -8,8 +8,10 @@ import { queueSeeds } from '~/shared/studio2/seed.mjs';
 import { readFileSync, writeFileSync, mkdirSync, existsSync, unlinkSync } from 'node:fs';
 import { join } from 'node:path';
 import template from '../../assets/studio2_minimax_native.json';
+import qwen21Template from '../../assets/studio2_qwen21_t2i.json';
 import { nativeVideoGraph, attachHeroReference, applyResolvedImageSize } from '~/shared/studio2/graphs.mjs';
 import { compilePrompt, scopedFile } from '~/shared/studio2/contracts.mjs';
+import { resolveQwen21Size } from '~/shared/studio2/input.mjs';
 import { createJob, restoreJob, getJob, emitJob, type Job } from '../jobs';
 import { markStudioLive, onLiveVideoSettled, type StudioJob } from '../studioQueue';
 import { readRecord, saveRecord, studio2Root, records } from './store';
@@ -74,8 +76,6 @@ async function prepareGraph(r: any) {
 
     let graph: any;
     if (['video', 'extend'].includes(q.mode)) {
-        if (q.engine !== 'minimax')
-            throw new Error('Studio 2 native identity video currently requires MiniMax. Use the existing xAIGen studio for LTX.');
         let start = a;
         if (q.startClipId) {
             const dir = join(studio2Root(), r.id);
@@ -86,6 +86,38 @@ async function prepareGraph(r: any) {
             start = await upload(r, 'start', readFileSync(dest));
             unlinkSync(dest);
         }
+        if (q.engine === 'ltx') {
+            const { buildWorkflow } = await import('../workflow');
+            const { frameLength } = await import('../videoChain');
+            const { ltxWorkflowEnabled, LTX_DISABLED_MESSAGE } = await import('~/utils/videoModels');
+            if (!ltxWorkflowEnabled()) throw new Error(LTX_DISABLED_MESSAGE);
+            const textToVideo = q.mode === 'video' && !start;
+            if (!start && !textToVideo) throw new Error('Choose a start still for LTX video.');
+            const fps = s.fps || 24;
+            graph = buildWorkflow({
+                workflow: textToVideo ? 'ltx-t2v' : 'ltx',
+                imageName: start || '',
+                prompt: q.compiledPrompt,
+                width: s.width,
+                height: s.height,
+                length: frameLength(s.duration || 5, fps),
+                fps,
+                steps: s.steps,
+                cfg: s.cfg,
+                seed: s.seed,
+                samplerName: 'euler',
+                scheduler: 'simple',
+                turbo: false,
+                duration: s.duration || 5,
+                filenamePrefix: prefix + '/video',
+                loraStack: s.loraStack,
+                useIdentityRefs: false
+            });
+            await ensureComfyLoraNames('video');
+            r.graphId = textToVideo ? 'workflow_ltx_video.json#t2v' : 'workflow_ltx_video.json';
+        } else {
+            if (q.engine !== 'minimax')
+                throw new Error('Studio 2 video supports MiniMax H3 and LTX (xAIGen).');
             const end = await load(q.endStillId, 'end'), guides = [];
             for (const [i, g] of q.guides.entries())
                 guides.push({ image: await load(g.stillId, `guide${i}`), frame: g.frame });
@@ -95,6 +127,29 @@ async function prepareGraph(r: any) {
             applyUserLoraToGraph(graph, s.loraStack);
             r.graphId = 'studio2_minimax_native.json';
         }
+    }
+    else if (q.engine === 'qwen21') {
+        if (q.mode !== 'generate')
+            throw new Error('Qwen 2.1 is T2I in this build.');
+        const size = resolveQwen21Size(s.aspect);
+        s.width = size.width;
+        s.height = size.height;
+        graph = structuredClone(qwen21Template);
+        graph['9'].inputs.prompt = q.compiledPrompt;
+        graph['9'].inputs.negative_prompt = stylePrompt(q.imageStyles, true);
+        // TextEncodeQwenImage21 builds the 64-ch empty latent from resolution (square T2I).
+        graph['9'].inputs.resolution = Math.max(size.width, size.height);
+        graph['15'].inputs.seed = s.seed;
+        graph['15'].inputs.steps = s.steps || 25;
+        graph['15'].inputs.cfg = s.cfg ?? 1;
+        graph['15'].inputs.sampler_name = 'euler';
+        graph['15'].inputs.scheduler = 'simple';
+        graph['21'].inputs.filename_prefix = prefix + '/image';
+        r.sampleLatent = 'qwen21 textencode latent';
+        r.sampleDenoise = null;
+        r.heroReferenceAttached = false;
+        r.graphId = 'studio2_qwen21_t2i.json';
+    }
     else {
         const mode = identityPlan?.mode ?? (q.mode === 'iterate' ? (a ? 'edit' : 'generate') : q.mode);
         const found = q.engine === 'krea' ? await resolveKreaGenerateAssets() : null;

server/utils/studio2/size.ts

@@ -21,5 +21,5 @@ export async function resolveRequestSize(owner: string, request: any) {
     }
   }
   if (video && request.startClipId && !dimensions?.width) throw new Error('Cannot measure the previous clip.')
-  Object.assign(request.settings,resolveSize(request.settings.aspect,dimensions))
+  Object.assign(request.settings,resolveSize(request.settings.aspect,dimensions,undefined,request.engine))
 }

server/utils/studioQueue.ts

@@ -13,7 +13,7 @@ import { imageV2StackSpecials } from '~/utils/imageV2'
 import { allowIdentityRefs, type PermanenceRef } from '~/utils/globalLocks'
 
 export type StudioJobStatus = 'waiting' | 'running' | 'held' | 'complete' | 'error' | 'cancelled'
-export type StudioJobKind = 'video' | 'edit' | 'music'
+export type StudioJobKind = 'video' | 'edit' | 'music' | 'caption'
 
 export interface StudioJobPayload {
   studio2Id?: string
@@ -40,6 +40,7 @@ export interface StudioJobPayload {
   useIdentityRefs: boolean
   stillId?: string
   stillFilename?: string
+  captionStyle?: string
   hideThumbnail: boolean
   hideInput?: boolean
   folderLocked?: boolean
@@ -224,6 +225,7 @@ export function listStudioJobs(owner: string) {
 export function studioJobKind(job: Pick<StudioJob, 'kind'> | { kind?: string }) {
   if (job.kind === 'edit') return 'edit'
   if (job.kind === 'music') return 'music'
+  if (job.kind === 'caption') return 'caption'
   return 'video'
 }
 
@@ -240,6 +242,8 @@ export function summarizeStudioJob(job: StudioJob) {
     familyId: job.familyId,
     shotQueueId: job.shotQueueId,
     liveJobId: job.liveJobId,
+    folderId: job.payload.folderId,
+    musicEngine: job.payload.musicEngine,
     stillId: job.payload.stillId,
     workflow: job.payload.workflow,
     imagePipeline: job.payload.imagePipeline || 'v1',
@@ -286,7 +290,7 @@ function failZombieLiveJob(job: Job, error: string) {
 function sweepStaleLiveJobs() {
   const now = Date.now()
   for (const job of listJobs()) {
-    if (job.studio2 || job.yueGp || job.yue2 || job.upscale) continue
+    if (job.studio2 || job.yueGp || job.yue2 || job.upscale || job.caption) continue
     if (job.saving) continue
     if (job.status === 'queued' && !job.promptId && now - job.startedAt >= QUEUED_GRACE_MS) {
       failZombieLiveJob(job, 'Job never started')
@@ -306,12 +310,12 @@ async function reapZombieLiveJobs() {
   const { fetchHistory } = await import('~/server/utils/comfy')
   for (const job of listJobs()) {
     // Never interrupt download/stitch/library save — Comfy is idle then by design.
-    if (job.studio2 || job.yueGp || job.yue2 || job.upscale) continue
+    if (job.studio2 || job.yueGp || job.yue2 || job.upscale || job.caption) continue
     if (job.saving) continue
     if (job.library?.chainContinuing) continue
     if (job.status !== 'running' && job.status !== 'uploading' && job.status !== 'queued') continue
     if (jobIsLocallySubmitting(job)) continue
-    const zombieMs = (job.kind === 'edit' || job.kind === 'music') ? 45_000 : ZOMBIE_EMPTY_COMFY_MS
+    const zombieMs = (job.kind === 'edit' || job.kind === 'music' || job.kind === 'caption') ? 45_000 : ZOMBIE_EMPTY_COMFY_MS
     if (jobAgeMs(job) < zombieMs) continue
     if (!job.promptId) {
       if (jobAgeMs(job) >= QUEUED_GRACE_MS) failZombieLiveJob(job, 'Job never started on ComfyUI')
@@ -351,7 +355,7 @@ async function reapZombieLiveJobs() {
 }
 
 function liveJobOwnsGpu(job: Job) {
-  if ((job.studio2 || job.yueGp || job.yue2 || job.upscale) && ['running', 'queued', 'uploading'].includes(job.status)) return true
+  if ((job.studio2 || job.yueGp || job.yue2 || job.upscale || job.caption) && ['running', 'queued', 'uploading'].includes(job.status)) return true
   if (job.library?.stopAfterCurrent) return false
   if (job.library?.chainContinuing) return true
   if (job.saving) return true
@@ -371,7 +375,7 @@ function liveJobOwnsGpu(job: Job) {
  */
 function clearDeadGpuClaimsForForceStart() {
   for (const live of listJobs()) {
-    if (live.studio2 || live.yueGp || live.yue2 || live.upscale) continue
+    if (live.studio2 || live.yueGp || live.yue2 || live.upscale || live.caption) continue
     if (live.saving) continue
     if (jobIsLocallySubmitting(live)) continue
     if (live.status !== 'running' && live.status !== 'queued' && live.status !== 'uploading') {
@@ -465,8 +469,11 @@ export async function addStudioJob(params: {
   kind?: StudioJobKind
 }) {
   const now = Date.now()
-  const kind = params.kind === 'edit' ? 'edit' : params.kind === 'music' ? 'music' : 'video'
-  const shotCount = kind === 'music'
+  const kind = params.kind === 'edit' ? 'edit'
+    : params.kind === 'music' ? 'music'
+      : params.kind === 'caption' ? 'caption'
+        : 'video'
+  const shotCount = kind === 'music' || kind === 'caption'
     ? 1
     : kind === 'edit'
       ? 1 + (params.payload.passes?.length || 0)
@@ -602,6 +609,10 @@ export async function clearStuckStudioWork(owner: string) {
       const { cancelYue2Job } = await import('./yue2')
       await cancelYue2Job(job)
     }
+    if (job.caption) {
+      const { cancelCaptionJob } = await import('./caption')
+      await cancelCaptionJob(job)
+    }
     job.status = 'cancelled'
     job.error = 'Cleared by force reset'
     if (job.library) {
@@ -674,6 +685,11 @@ async function stopLiveGeneration(liveJobId?: string, shotQueueId?: string) {
     await cancelYue2Job(live)
     return
   }
+  if (live?.caption) {
+    const { cancelCaptionJob } = await import('./caption')
+    await cancelCaptionJob(live)
+    return
+  }
   if (live) {
     live.status = 'cancelled'
     if (live.library) {
@@ -881,6 +897,14 @@ function pendingAlive(job: StudioJob) {
   if (job.liveJobId) {
     const pending = readPendingJob(job.liveJobId)
     if (pending?.promptId && freshEnough(pending.startedAt)) return true
+    if (job.kind === 'music') {
+      const root = join(String(useRuntimeConfig().libraryDir || process.env.LIBRARY_DIR || '/data/library'), 'yue2-pending', `${job.liveJobId}.json`)
+      if (existsSync(root)) return true
+    }
+    if (job.kind === 'caption') {
+      const root = join(String(useRuntimeConfig().libraryDir || process.env.LIBRARY_DIR || '/data/library'), 'caption-pending', `${job.liveJobId}.json`)
+      if (existsSync(root)) return true
+    }
   }
   if (!job.shotQueueId) return false
   return listPendingJobs().some(pending => (
@@ -923,7 +947,7 @@ function repairStaleJobs(jobs: StudioJob[]) {
       job.updatedAt = Date.now()
       continue
     }
-    if (!job.payload.upscale && job.payload.musicEngine !== 'yue' && job.payload.musicEngine !== 'yue2' && job.status === 'error' && isTransientComfyError(job.lastError)) {
+    if (!job.payload.upscale && job.kind !== 'music' && job.status === 'error' && isTransientComfyError(job.lastError)) {
       job.status = 'waiting'
       job.liveJobId = undefined
       job.lastError = undefined
@@ -932,7 +956,7 @@ function repairStaleJobs(jobs: StudioJob[]) {
       job.updatedAt = Date.now()
       continue
     }
-    if (!job.payload.upscale && job.payload.musicEngine !== 'yue' && job.payload.musicEngine !== 'yue2' && job.status === 'held' && isTransientComfyError(job.lastError)) {
+    if (!job.payload.upscale && job.kind !== 'music' && job.status === 'held' && isTransientComfyError(job.lastError)) {
       job.status = 'waiting'
       job.liveJobId = undefined
       job.lastError = undefined
@@ -1389,6 +1413,42 @@ async function startStudioExtendJob(item: StudioJob) {
   }
 }
 
+async function startStudioCaptionJob(item: StudioJob) {
+  let live: import('~/server/utils/jobs').Job | undefined
+  try {
+    const { startCaptionJob } = await import('./caption')
+    const style = String(item.payload.captionStyle || 'descriptive')
+    const stillId = String(item.payload.stillId || '')
+    if (!stillId) throw new Error('Caption requires a still.')
+    live = await startCaptionJob({
+      ownerKey: item.ownerKey,
+      folderId: item.payload.folderId,
+      stillId,
+      captionStyle: style,
+      studio2Id: item.payload.studio2Id,
+      name: item.payload.name
+    })
+    await markStudioLive(item.ownerKey, item.id, live.id)
+    if (item.payload.studio2Id) {
+      const { readRecord, saveRecord } = await import('./studio2/store')
+      const record = readRecord(item.ownerKey, item.payload.studio2Id)
+      record.state = 'rendering'
+      record.liveId = live.id
+      record.startedAt = Date.now()
+      record.updatedAt = Date.now()
+      saveRecord(record)
+    }
+  } catch (error) {
+    const message = error instanceof Error ? error.message : String(error)
+    if (live && (live.status === 'queued' || live.status === 'uploading' || live.status === 'running')) {
+      live.status = 'error'
+      live.error = message
+    }
+    await parkStudioOnStartFailure(item.ownerKey, item.id, message)
+    kickStudioQueue()
+  }
+}
+
 async function startStudioMusicJob(item: StudioJob) {
   let live: import('~/server/utils/jobs').Job | undefined
   try {
@@ -1407,8 +1467,7 @@ async function startStudioMusicJob(item: StudioJob) {
       lyricsStrength: payload.instrumental ? 0 : (payload.lyricsStrength ?? 0.9),
       instrumental: payload.instrumental === true,
       folderLocked: payload.folderLocked,
-      engine: payload.musicEngine || 'ace-step',
-      yueProfile: payload.yueProfile ?? 1,
+      engine: 'yue2',
       samplerName: payload.samplerName || 'euler',
       scheduler: payload.scheduler || 'simple'
     })
@@ -1432,6 +1491,10 @@ export async function startStudioJob(item: StudioJob) {
 }
 
 async function startStudioJobReserved(item: StudioJob) {
+  if (studioJobKind(item) === 'caption') {
+    await startStudioCaptionJob(item)
+    return
+  }
   if (item.payload.studio2Id) {
     try {
       const { startStudio2Job } = await import('./studio2/runner')
@@ -1696,7 +1759,7 @@ export async function onLiveVideoSettled(job: Job) {
     return
   }
   const remaining = remainingStudioShots(job)
-  const wakeFail = !job.studio2 && !job.yueGp && !job.yue2 && !job.upscale && job.status === 'error' && remaining > 0 && isTransientComfyError(job.error)
+  const wakeFail = !job.studio2 && !job.yueGp && !job.yue2 && !job.upscale && !job.caption && job.status === 'error' && remaining > 0 && isTransientComfyError(job.error)
   const failed = (job.status === 'error' || job.status === 'cancelled') && !wakeFail
 
   await mutateStore(owner, (store) => {

server/utils/yue2.ts

@@ -11,6 +11,11 @@ function settings() {
     token: String(config.comfyControlToken || process.env.COMFY_CONTROL_TOKEN || '') }
 }
 export function yue2Configured() { return Boolean(settings().url) }
+function pendingRoot() { return join(String(useRuntimeConfig().libraryDir || process.env.LIBRARY_DIR || '/data/library'), 'yue2-pending') }
+export function yue2PendingAlive(jobId: string) {
+  if (!jobId) return false
+  return existsSync(join(pendingRoot(), `${jobId}.json`))
+}
 async function request(path: string, body?: unknown) {
   const { url, token } = settings()
   if (!url) throw new Error('YuE2 host is not configured. Set COMFY_CONTROL_URL.')
@@ -26,7 +31,6 @@ async function request(path: string, body?: unknown) {
   }
   return response
 }
-function pendingRoot() { return join(String(useRuntimeConfig().libraryDir || process.env.LIBRARY_DIR || '/data/library'), 'yue2-pending') }
 function persist(job: Job) {
   mkdirSync(pendingRoot(), { recursive: true })
   const path = join(pendingRoot(), `${job.id}.json`)

shared/studio2/caption.mjs

@@ -0,0 +1,66 @@
+export const CAPTION_STYLES = ['descriptive', 'klein_prompt', 'delta', 'tags']
+
+export const DELTA_LOCKED_LOOK =
+  'short light-brown bob, pink tank, light blue shorts, open green Slytherin robe with red-gold trim, barefoot'
+
+export function captionStylePrompt(style) {
+  switch (style) {
+    case 'klein_prompt':
+      return 'Write one self-contained image generation prompt for Flux Klein. Include subject, pose, clothing, setting, lighting, and camera. No preamble or quotes.'
+    case 'delta':
+      return `The character's locked baseline look is: ${DELTA_LOCKED_LOOK}. Describe ONLY what changed from that baseline in this image. Do not restate the locked look.`
+    case 'tags':
+      return 'List comma-separated descriptive tags for this image. Tags only, no sentences.'
+    case 'descriptive':
+    default:
+      return 'Describe this image in clear, detailed prose. Cover subject, appearance, clothing, pose, setting, and lighting.'
+  }
+}
+
+export function validateCaptionRequest(raw) {
+  const fail = message => { throw Object.assign(new Error(message), { statusCode: 400 }) }
+  const body = raw || {}
+  const style = String(body.captionStyle || body.style || 'descriptive')
+  if (!CAPTION_STYLES.includes(style)) fail('Unknown caption style.')
+  const folderId = String(body.folderId || '')
+  if (!folderId) fail('Choose a folder.')
+  const stillId = String(body.stillId || body.imageAId || '')
+  const sourcePath = String(body.sourcePath || '')
+  if (!stillId && !sourcePath) fail('Choose a still to describe.')
+  return {
+    mode: 'caption',
+    folderId,
+    stillId,
+    sourcePath,
+    captionStyle: style,
+    promptSections: { action: `Describe · ${style}` },
+    compiledPrompt: captionStylePrompt(style),
+    engine: 'caption',
+    imageStyles: { positive: [], negative: [] },
+    hideOutput: false,
+    lockFace: false,
+    lockOutfit: false,
+    hideInput: false,
+    identityStillId: '',
+    imageAId: stillId,
+    imageBId: '',
+    maskId: '',
+    startClipId: '',
+    endStillId: '',
+    guides: [],
+    batch: [],
+    settings: {
+      aspect: 'auto',
+      fps: 24,
+      width: 0,
+      height: 0,
+      steps: 1,
+      cfg: 0,
+      seedMode: 'random',
+      seed: 0,
+      duration: 0,
+      turbo: true,
+      loraStack: []
+    }
+  }
+}

shared/studio2/contracts.mjs

@@ -13,7 +13,7 @@ export function appearanceLock({lockFace = true, lockOutfit = true} = {}) {
 export function compilePrompt(value, video = false, locks = {}) {
   if (video) return [compileSections(value),appearanceLock(locks)].filter(Boolean).join('\n\n')
   const hasImage=!!(locks.imageAId || locks.identityStillId)
-  const attached=hasImage && locks.engine !== 'krea'
+  const attached=hasImage && locks.engine !== 'krea' && locks.engine !== 'qwen21'
   const reference=locks.mode!=='generate' && locks.identityStillId && locks.imageAId && locks.identityStillId!==locks.imageAId ? 'second (hero) reference' : 'image reference'
   return [sectionsFrom(value).action, stylePrompt(locks.imageStyles),
     attached && locks.lockFace !== false ? `Preserve the person’s face, glasses, hair, and body identity from the ${reference}.` : '',
@@ -31,9 +31,10 @@ export function validateRequest(raw, xaigen = false) {
   const mode = raw.mode || 'generate', engine = raw.engine || 'flux'
   if (!['generate','edit','compose','refine','iterate','video','extend'].includes(mode)) fail('Unknown task.')
   const video = ['video','extend'].includes(mode)
-  if (!(video ? ['minimax','ltx'] : ['flux','krea']).includes(engine)) fail('Engine does not support this task.')
+  if (!(video ? ['minimax','ltx'] : ['flux','krea','qwen21']).includes(engine)) fail('Engine does not support this task.')
   if (engine === 'ltx' && !xaigen) fail('LTX is available only in xAIGen.')
   if (engine === 'krea' && ['edit','compose','iterate'].includes(mode)) fail('Sticky hero edits require Klein. Krea remains available for Generate and masked Refine; no engine fallback is performed.')
+  if (engine === 'qwen21' && mode !== 'generate') fail('Qwen 2.1 is T2I in this build.')
   if (Array.isArray(raw.batch) && raw.batch.length>30) fail('A batch supports up to 31 shots including the first prompt.')
   if (Array.isArray(raw.guides) && raw.guides.length>8) fail('Use up to eight guide stills.')
   if (!raw.folderId) fail('Choose a folder.')
@@ -46,20 +47,24 @@ export function validateRequest(raw, xaigen = false) {
   const promptSections = video ? allSections : {action:allSections.action}
   if (!promptSections.action.trim()) fail('Enter an action prompt.')
   const number=(key, fallback, min, max) => { const n=Number(raw.settings?.[key] ?? fallback); if (!Number.isFinite(n)||n<min||n>max) fail(`Invalid ${key}.`); return n }
-  for (const key of ['steps']) if (!Number.isInteger(number(key,key==='steps'?24:1,0,2147483647))) fail(`Invalid ${key}.`)
-  for (const key of ['width','height']) if (number(key,key==='width'?1344:768,256,4096)%32) fail('Dimensions must be multiples of 32.')
+  const qwen = engine === 'qwen21'
+  const defaultSteps = video ? 8 : qwen ? 25 : 24
+  const defaultCfg = video ? 1.5 : qwen ? 1 : 4
+  const defaultWidth = qwen ? 1024 : 1344
+  const defaultHeight = qwen ? 1024 : 768
+  for (const key of ['steps']) if (!Number.isInteger(number(key,key==='steps'?defaultSteps:1,0,2147483647))) fail(`Invalid ${key}.`)
+  for (const key of ['width','height']) if (number(key,key==='width'?defaultWidth:defaultHeight,256,4096)%32) fail('Dimensions must be multiples of 32.')
   const duration=number('duration',5,1,15)
   if (video) {
     if (promptSections.cameraBeats.slice(0,-1).some(b=>b.durationSec == null)) fail('Only the last camera beat can run to the end.')
     if (promptSections.cameraBeats.reduce((sum,b)=>sum+(b.durationSec || 0),0)>duration) fail('Camera beats exceed the clip duration.')
-    if (raw.lockFace === false && raw.lockOutfit === false && !raw.imageAId && !raw.startClipId) fail('Choose a start still or previous clip when both identity locks are off.')
   }
   const frames=Math.ceil((duration*number('fps',24,1,60)-5)/17)*17+5
   const shotSeeds=queueSeeds(raw.settings,1+(Array.isArray(raw.batch)?raw.batch.length:0))
-  return { mode, engine, shotSeeds, imageStyles:video?{positive:[],negative:[]}:normalizeStyles(raw.imageStyles), hideOutput:!!raw.hideOutput, lockFace:raw.lockFace !== false, lockOutfit:raw.lockOutfit !== false, hideInput:!!raw.hideInput, folderId:String(raw.folderId), identityStillId:raw.identityStillId || '', imageAId:raw.imageAId || '', imageBId:raw.imageBId || '', maskId:raw.maskId || '', startClipId:raw.startClipId || '', endStillId:raw.endStillId || '',
+  return { mode, engine, shotSeeds, imageStyles:video?{positive:[],negative:[]}:normalizeStyles(raw.imageStyles), hideOutput:!!raw.hideOutput, lockFace:qwen ? false : raw.lockFace !== false, lockOutfit:qwen ? false : raw.lockOutfit !== false, hideInput:!!raw.hideInput, folderId:String(raw.folderId), identityStillId:qwen ? '' : (raw.identityStillId || ''), imageAId:raw.imageAId || '', imageBId:raw.imageBId || '', maskId:raw.maskId || '', startClipId:raw.startClipId || '', endStillId:raw.endStillId || '',
     startFrameSource: raw.startClipId ? {kind:'previous-last-frame',clipId:raw.startClipId} : {kind:'still',stillId:raw.imageAId || ''},
-    promptSections, compiledPrompt:compilePrompt(promptSections,video,raw),
-    settings: { aspect:['auto','16:9','9:16','1:1'].includes(raw.settings?.aspect)?raw.settings.aspect:'auto', fps:number('fps',24,1,60), width:number('width',1344,256,4096),height:number('height',768,256,4096),steps:number('steps',video?8:24,1,100),cfg:number('cfg',video?1.5:4,0,20),seedMode:raw.settings?.seedMode==='locked'?'locked':'random',seed:shotSeeds[0],duration:number('duration',5,1,15),turbo:raw.settings?.turbo !== false, loraStack:Array.isArray(raw.settings?.loraStack)?raw.settings.loraStack:[] },
+    promptSections, compiledPrompt:compilePrompt(promptSections,video,{...raw,engine,lockFace:qwen?false:raw.lockFace,lockOutfit:qwen?false:raw.lockOutfit}),
+    settings: { aspect:['auto','16:9','9:16','1:1'].includes(raw.settings?.aspect)?raw.settings.aspect:'auto', fps:number('fps',24,1,60), width:number('width',defaultWidth,256,4096),height:number('height',defaultHeight,256,4096),steps:number('steps',defaultSteps,1,100),cfg:number('cfg',defaultCfg,0,20),seedMode:raw.settings?.seedMode==='locked'?'locked':'random',seed:shotSeeds[0],duration:number('duration',5,1,15),turbo:raw.settings?.turbo !== false, loraStack:qwen?[]:(Array.isArray(raw.settings?.loraStack)?raw.settings.loraStack:[]) },
     guides: (Array.isArray(raw.guides)?raw.guides:[]).slice(0,8).map(g=>{ if (!g.stillId || !Number.isInteger(g.frame) || g.frame<1 || g.frame>=frames) fail('Each guide requires a still and positive frame index.'); return {stillId:String(g.stillId),frame:g.frame} }),
     batch: (Array.isArray(raw.batch)?raw.batch:[]).slice(0,30).map(s=>{const next=sectionsFrom({...promptSections,...(typeof s==='string'?{action:s}:s)});if(!next.action.trim())fail('Every shot needs an action prompt.');return video?next:{action:next.action}}) }
 }

shared/studio2/graphs.mjs

@@ -2,9 +2,20 @@ export function nativeVideoGraph(template, request, files, prefix) {
   const appearance = request.lockFace !== false || request.lockOutfit !== false
   if (appearance && !files.hero) throw new Error('Identity still is required; never substitute the start frame.')
   const g=structuredClone(template), s=request.settings
+  const textToVideo = request.mode === 'video' && !appearance && !files.start
   if (appearance) { g.hero.inputs.image=files.hero; g['155'].inputs['ref_images.ref_image_0']=['hero',0] }
   else { delete g.hero; delete g['155'].inputs['ref_images.ref_image_0'] }
   Object.assign(g['155'].inputs,{prompt:request.compiledPrompt,width:s.width,height:s.height,length:Math.ceil((s.duration*(s.fps || 24)-5)/17)*17+5,ref_image_size:'max'})
+  if (textToVideo) {
+    // Studio 1 T2V: MiniMaxH3ImageToVideo with first_frame disconnected (no empty placeholder still).
+    g['155'].class_type = 'MiniMaxH3ImageToVideo'
+    g['155']._meta = { title: 'MiniMax H3 Text to Video' }
+    delete g['155'].inputs['ref_images.ref_image_0']
+    delete g['155'].inputs.ref_image_size
+    delete g['155'].inputs.audio_vae
+    delete g['155'].inputs.first_frame
+    delete g['155'].inputs.last_frame
+  }
   let conditioning=['155',0]
   for (const [name,image,frame] of [['start',files.start,0],['end',files.end,-1],...(files.guides||[]).map((f,i)=>[`guide${i}`,f.image,f.frame])]) {
     if (!image) continue

shared/studio2/input.mjs

@@ -1,7 +1,13 @@
 import {restoreLoras,loraUse} from './lora-memory.mjs'
 import { restoreInput } from './contracts.mjs'
 export const aspectOptions = [{id:'auto',label:'Automatic',hint:'Match start or hero still'},{id:'16:9',label:'16:9',hint:'1344 × 768 · padded'},{id:'9:16',label:'9:16',hint:'768 × 1344 · padded'},{id:'1:1',label:'1:1',hint:'960 × 960 · padded'}]
-export function resolveSize(aspect, source, fallback = {width:1344,height:768}) {
+export function resolveQwen21Size(aspect) {
+  const presets = {'16:9':[1024,576],'9:16':[576,1024],'1:1':[1024,1024]}
+  if (presets[aspect]) return {width:presets[aspect][0],height:presets[aspect][1]}
+  return {width:1024,height:1024}
+}
+export function resolveSize(aspect, source, fallback = {width:1344,height:768}, engine = '') {
+  if (engine === 'qwen21') return resolveQwen21Size(aspect)
   const presets = {'16:9':[1344,768],'9:16':[768,1344],'1:1':[960,960]}
   if (presets[aspect]) return {width:presets[aspect][0],height:presets[aspect][1]}
   if (!source?.width || !source?.height) return {width:fallback.width || 1344,height:fallback.height || 768}
@@ -35,7 +41,7 @@ export function restoreAsset(form, item, xaigen = false, memory) {
   const restored = restoreInput(item)
   const next = {...form,...restored,settings:{...form.settings,...restored.settings}}
   next.settings.seedMode='random'
-  next.engine=restored.engine==='krea'?'krea':'flux'
+  next.engine=restored.engine==='qwen21'?'qwen21':restored.engine==='krea'?'krea':'flux'
   next.mode=['generate','edit','compose','refine','iterate'].includes(restored.mode)?restored.mode:'edit'
   next.imageAId=item.id
   next.startClipId=''
@@ -62,7 +68,7 @@ export function restoreAll(form,item,xaigen=false,memory) {
  const restored=restoreInput(item),next={...form,...restored,settings:{...form.settings,...restored.settings,seedMode:'random'}}
  if(item.kind==='clip') return {...bindClip(next,item,xaigen,memory),mode:'extend'}
  next.mode=restored.mode==='iterate'?'edit':restored.mode || 'edit'
- next.engine=restored.engine==='krea'?'krea':'flux'
+ next.engine=restored.engine==='qwen21'?'qwen21':restored.engine==='krea'?'krea':'flux'
  next.imageAId=item.id
  next.startClipId=''
  return restoreLoras(next,item.studio2?.settings?.loraStack ?? item.settings?.loraStack,xaigen,memory)

shared/studio2/lora-cache.mjs

@@ -1,10 +1,13 @@
-export async function cachedLoras({read,write,live,fresh=false}) {
+export async function cachedLoras({read,write,live,fresh: _fresh=false}) {
  const cached=read() || {image:[],video:[]};
- if (!fresh && (cached.image?.length || cached.video?.length)) return cached;
  try {
+  // Always live-scan. Disk is asleep/empty fallback only — never a stale snapshot that skips Comfy.
   const listed=await live();
-  const next={image:listed.image?.length?listed.image:cached.image||[],video:listed.video?.length?listed.video:cached.video||[]};
-  if(next.image.length || next.video.length) write(next);
+  const image=Array.isArray(listed?.image)?listed.image:[];
+  const video=Array.isArray(listed?.video)?listed.video:[];
+  if (!image.length && !video.length) return cached;
+  const next={image,video};
+  write(next);
   return next;
  } catch { return cached; }
 }

shared/studio2/lora-memory.mjs

@@ -2,7 +2,9 @@ import {filterLoraStackForUse} from '../../utils/loras.ts'
 
 export function loraUse(form) {
  const video=['video','extend'].includes(form.mode)
- return video ? (form.engine==='ltx'?'ltx':'minimax') : (form.engine==='krea'?'krea':'flux')
+ if (video) return form.engine==='ltx'?'ltx':'minimax'
+ if (form.engine==='qwen21') return 'qwen21'
+ return form.engine==='krea'?'krea':'flux'
 }
 
 // One controller per bench. Storage is optional (SSR/private browsing still works).

shared/studio2/status.mjs

@@ -1,7 +1,16 @@
-export function gpuStatus(jobs, health, upscales=[]) {
+export function gpuStatus(jobs, health, upscales=[], musicJobs=[], captionJobs=[]) {
  if (!health) return {label:'Status unavailable',tone:'unknown',detail:'Checking GPU connection…'}
  const upscale=upscales.find(j=>j.status==='running')
  if(upscale)return {label:'Upscaling',tone:'busy',detail:upscale.name || 'Video upscale'}
+ const music=musicJobs.find(j=>['waiting','running','held'].includes(j.status) || j.live?.status==='running' || j.live?.musicActivity?.running)
+ if(music){
+  const stage=music.live?.message || music.live?.stage || music.waitReason || 'YuE2'
+  return {label:music.status==='waiting'?'Queued':'Generating',tone:'busy',detail:String(stage)}
+ }
+ const caption=captionJobs.find(j=>['waiting','waking','submitting','rendering','saving'].includes(j.state) || j.state==='queued' || j.state==='held')
+ if(caption)return {label:caption.state==='waiting'?'Queued':'Describing',tone:'busy',detail:caption.request?.captionStyle || 'Caption'}
+ if(health.caption?.busy)return {label:'Describing',tone:'busy',detail:'Caption model is using the GPU'}
+ if(health.yue2?.busy)return {label:'Generating',tone:'busy',detail:'YuE2 is using the GPU'}
  const job=jobs.find(j=>['waking','submitting','rendering','saving'].includes(j.state))
  if(job)return {label:{waking:'Waking up',submitting:'Submitting',rendering:'Generating',saving:'Saving'}[job.state],tone:'busy',detail:job.request?.promptSections?.action || 'Generation in progress'}
  if(health.comfy?.busy || health.imageComfy?.busy)return {label:'Busy',tone:'busy',detail:'GPU is working on another job'}

shared/studio2/viewer.mjs

@@ -1,12 +1,22 @@
 export function folderAssets(library, jobs, folderId) {
  const metadata=new Map(jobs.flatMap(j=>j.outputs || []).map(a=>[a.id,a.studio2]))
- return ['still','clip'].flatMap(kind=>(library[kind==='still'?'stills':'clips'] || []).filter(a=>a.id && a.folderId===folderId && a.hasVideo!==false && a.available!==false).map(a=>({...a,kind,studio2:a.studio2 || metadata.get(a.id)}))).sort((a,b)=>(a.createdAt || a.studio2?.savedAt || 0)-(b.createdAt || b.studio2?.savedAt || 0))
+ const kinds=[['still','stills'],['clip','clips'],['track','tracks']]
+ return kinds.flatMap(([kind,key])=>(library[key] || []).filter(a=>a.id && a.folderId===folderId && a.hasVideo!==false && a.available!==false).map(a=>({...a,kind,studio2:a.studio2 || metadata.get(a.id)}))).sort((a,b)=>(a.createdAt || a.studio2?.savedAt || 0)-(b.createdAt || b.studio2?.savedAt || 0))
 }
 export function newestOutput(jobs, seen, folderId) {
  return jobs.filter(j=>j.request.folderId===folderId).flatMap(j=>j.outputs || []).filter(a=>!seen.has(a.id)).sort((a,b)=>(a.studio2?.savedAt || 0)-(b.studio2?.savedAt || 0)).at(-1) || null
 }
 export function reconcileSelection(selected, assets) { return assets.find(a=>a.id===selected?.id) || assets.at(-1) || null }
 export function activeJobs(jobs) { return jobs.filter(j=>['waking','submitting','rendering','saving'].includes(j.state)) }
-export function previewHidden(asset, xaigen, hideInput, hideOutput) { return !!xaigen && (asset?.role==='input' ? !!hideInput : !!hideOutput) }
+export function previewHidden(asset, revealed, xaigen) {
+ if (!xaigen || !asset?.id) return false
+ return !revealed?.has?.(asset.id)
+}
 export function selectResult(state, asset) { return {...state,selected:asset} }
-export function assetLabel(asset) { return asset?.name || asset?.originalName || asset?.filename || (asset ? (asset.kind==='clip'?'Clip ':'Still ')+asset.id.slice(0,8) : 'None') }
+export function assetLabel(asset) {
+ if (!asset) return 'None'
+ if (asset.name || asset.originalName || asset.filename) return asset.name || asset.originalName || asset.filename
+ if (asset.kind==='track') return 'Track '+asset.id.slice(0,8)
+ if (asset.kind==='clip') return 'Clip '+asset.id.slice(0,8)
+ return 'Still '+asset.id.slice(0,8)
+}

tests/caption.test.mjs

@@ -0,0 +1,139 @@
+import test from 'node:test'
+import assert from 'node:assert/strict'
+import { EventEmitter } from 'node:events'
+import { PassThrough } from 'node:stream'
+import { mkdtempSync, writeFileSync, readFileSync, mkdirSync, existsSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { createCaptionHost, validateCaptionHostRequest, captionModelFiles } from '../scripts/caption-host.mjs'
+import { validateCaptionRequest, captionStylePrompt, CAPTION_STYLES, DELTA_LOCKED_LOOK } from '../shared/studio2/caption.mjs'
+
+test('caption styles and delta lock text', () => {
+  assert.deepEqual(CAPTION_STYLES, ['descriptive', 'klein_prompt', 'delta', 'tags'])
+  assert.match(captionStylePrompt('delta'), new RegExp(DELTA_LOCKED_LOOK))
+  assert.match(captionStylePrompt('klein_prompt'), /Flux Klein/)
+  assert.match(captionStylePrompt('tags'), /comma-separated/)
+})
+
+test('validateCaptionRequest requires folder + still', () => {
+  assert.throws(() => validateCaptionRequest({}), /folder/)
+  assert.throws(() => validateCaptionRequest({ folderId: 'f1' }), /still/)
+  assert.throws(() => validateCaptionRequest({ folderId: 'f1', stillId: 's1', captionStyle: 'nope' }), /style/)
+  const ok = validateCaptionRequest({ folderId: 'f1', stillId: 's1', captionStyle: 'descriptive' })
+  assert.equal(ok.mode, 'caption')
+  assert.equal(ok.captionStyle, 'descriptive')
+  assert.equal(ok.lockFace, false)
+  assert.equal(ok.lockOutfit, false)
+})
+
+test('host validates job body', () => {
+  assert.throws(() => validateCaptionHostRequest({ id: 'bad', style: 'descriptive' }))
+  assert.deepEqual(
+    validateCaptionHostRequest({ id: 'caption-job-123456', style: 'tags' }),
+    { id: 'caption-job-123456', style: 'tags', imagePath: '' }
+  )
+})
+
+function fixture() {
+  const root = mkdtempSync(join(tmpdir(), 'aigen-caption-'))
+  const models = join(root, 'models')
+  mkdirSync(models)
+  writeFileSync(join(models, captionModelFiles.model), 'model')
+  writeFileSync(join(models, captionModelFiles.mmproj), 'mmproj')
+  const llama = join(root, 'llama-server.exe')
+  writeFileSync(llama, '')
+  const dataDir = join(root, 'jobs')
+  const child = new EventEmitter()
+  child.pid = 4242
+  child.stdout = new PassThrough()
+  child.stderr = new PassThrough()
+  child.kill = () => { setImmediate(() => child.emit('close', 0)); return true }
+  child.exitCode = null
+  const calls = []
+  let fetchCount = 0
+  const fetchImpl = async (url, init) => {
+    fetchCount++
+    if (String(url).endsWith('/health')) return { ok: true }
+    if (String(url).includes('/v1/chat/completions')) {
+      assert.match(JSON.stringify(JSON.parse(init.body)), /Describe this image|Flux Klein|locked baseline|comma-separated/)
+      return {
+        ok: true,
+        json: async () => ({ choices: [{ message: { content: 'a warm caption' } }] })
+      }
+    }
+    throw new Error(`unexpected fetch ${url}`)
+  }
+  const host = createCaptionHost({
+    modelsDir: models,
+    llamaServer: llama,
+    dataDir,
+    prepare: async () => {},
+    leaseValid: () => true,
+    spawnProcess: (...args) => { calls.push(args); setImmediate(() => {}); return child },
+    fetchImpl
+  })
+  return { host, child, calls, root, dataDir, fetchCount: () => fetchCount }
+}
+
+test('host loads llama-server, captions once, and unloads', async () => {
+  const f = fixture()
+  const jobDir = join(f.dataDir, 'caption-job-123456')
+  mkdirSync(jobDir, { recursive: true })
+  writeFileSync(join(jobDir, 'input.png'), Buffer.alloc(64, 1))
+  const started = f.host.start({ id: 'caption-job-123456', style: 'descriptive' }, 'lease')
+  // Let the host reach waitForServer / chat completion.
+  await new Promise(r => setTimeout(r, 50))
+  f.child.emit('spawn')
+  const result = await started
+  assert.equal(result.status, 'complete')
+  assert.equal(result.text, 'a warm caption')
+  assert.equal(f.calls.length, 1)
+  const [, args] = f.calls[0]
+  assert.ok(args.includes('--mmproj'))
+  assert.ok(args.includes('-m'))
+  assert.equal(f.host.read('caption-job-123456').text, 'a warm caption')
+  // busy may briefly remain true during the post-unload hold window
+  assert.ok(typeof f.host.busy() === 'boolean')
+})
+
+test('host rejects a second job while busy', async () => {
+  const f = fixture()
+  const dataDir = join(f.root, 'jobs2')
+  let resolveHealth
+  const healthGate = new Promise(resolve => { resolveHealth = resolve })
+  const hung = createCaptionHost({
+    modelsDir: join(f.root, 'models'),
+    llamaServer: join(f.root, 'llama-server.exe'),
+    dataDir,
+    prepare: async () => {},
+    leaseValid: () => true,
+    spawnProcess: () => {
+      const child = new EventEmitter()
+      child.pid = 99
+      child.stdout = new PassThrough()
+      child.stderr = new PassThrough()
+      child.kill = () => { setImmediate(() => child.emit('close', 1)); return true }
+      child.exitCode = null
+      return child
+    },
+    fetchImpl: async (url) => {
+      if (String(url).endsWith('/health')) {
+        await healthGate
+        return { ok: true }
+      }
+      return { ok: true, json: async () => ({ choices: [{ message: { content: 'x' } }] }) }
+    }
+  })
+  mkdirSync(join(dataDir, 'caption-job-aaaaaa'), { recursive: true })
+  writeFileSync(join(dataDir, 'caption-job-aaaaaa', 'input.png'), Buffer.alloc(64, 1))
+  const first = hung.start({ id: 'caption-job-aaaaaa', style: 'tags' }, 'lease')
+  await new Promise(r => setTimeout(r, 20))
+  await assert.rejects(hung.start({ id: 'caption-job-bbbbbb', style: 'tags' }, 'lease'), /already running/)
+  await hung.cancel('caption-job-aaaaaa')
+  resolveHealth()
+  await assert.rejects(first)
+})
+
+test('POST /caption sync shape is documented on host module', () => {
+  assert.equal(typeof createCaptionHost({ prepare: async () => {}, leaseValid: () => true }).captionOnce, 'function')
+})

tests/studio2-bench.test.mjs

@@ -1,9 +1,6 @@
 import test from 'node:test'
 import assert from 'node:assert/strict'
 import {readFileSync} from 'node:fs'
-import * as Vue from 'vue'
-import {compile} from '@vue/compiler-dom'
-import {renderToString} from '@vue/server-renderer'
 import ts from 'typescript'
 import {folderAssets,reconcileSelection,activeJobs,previewHidden} from '../shared/studio2/viewer.mjs'
 import {useAsHero,useAsSource,restoreAll} from '../shared/studio2/input.mjs'
@@ -24,19 +21,45 @@ test('bench includes only active execution phases and clear finished retains int
  assert.equal(fn({}).cleared,1);assert.equal(rows[0].dismissed,true);assert.equal(rows[0].cleanupPending.length,1);assert.equal(rows[1].dismissed,undefined)
 })
 test('hidden xAIGen canvas and family render no image or video elements',async()=>{
+ const revealed=new Set()
+ assert.equal(previewHidden({id:'a',role:'input'},revealed,true),true)
+ assert.equal(previewHidden({id:'a',role:'input'},new Set(['a']),true),false)
+ assert.equal(previewHidden({id:'a'},revealed,false),false)
  const src=readFileSync(new URL('../components/studio2/Bench.vue',import.meta.url),'utf8')
- const stage=src.slice(src.indexOf('    <div class="flex min-h-80'),src.indexOf('    <div v-if="selected"'))
- const strip=src.split('\n').find(line=>line.includes('aria-label="Current family outputs"'))
- const render=new Function('Vue',compile('<section>'+stage+strip+'</section>',{mode:'function',prefixIdentifiers:true}).code)(Vue)
- const state={selected:{id:'a',kind:'still'},hideOutput:true,xaigen:true,family:[{id:'a',kind:'still'}],form:{hideInput:true},previewHidden,stillUrl:id=>'/image/'+id,selectOutput(){}}
- const hidden=await renderToString(Vue.createSSRApp({render,data:()=>state}));assert.doesNotMatch(hidden,/<img|<video/);assert.match(hidden,/preview hidden/)
- const visible=await renderToString(Vue.createSSRApp({render,data:()=>({...state,xaigen:false})}));assert.match(visible,/<img/)
- assert.equal(previewHidden({role:'input'},true,true,false),true)
+ assert.match(src,/aria-label="Current family outputs"/)
+ assert.match(src,/preview hidden/)
+ assert.doesNotMatch(src,/Hide input stills/)
+ assert.doesNotMatch(src,/Hide output preview/)
 })
 test('cached LoRAs survive sleeping GPU and failed explicit refresh',async()=>{
  const cache={image:['Klein/portrait.safetensors'],video:['motion.safetensors']};let queried=0
  const opts={read:()=>cache,write:()=>assert.fail('Do not erase cache'),live:async()=>{queried++;throw Error('Comfy asleep')}}
- assert.deepEqual(await cachedLoras(opts),cache);assert.equal(queried,0);assert.deepEqual(await cachedLoras({...opts,fresh:true}),cache);assert.equal(queried,1)
+ assert.deepEqual(await cachedLoras(opts),cache);assert.equal(queried,1)
+ assert.deepEqual(await cachedLoras({...opts,fresh:true}),cache);assert.equal(queried,2)
+})
+test('bench load and Refresh always live-scan and replace disk so new Comfy names appear',async()=>{
+ let store={image:['old.safetensors'],video:['old-vid.safetensors']}
+ const next=await cachedLoras({
+  read:()=>store,
+  write:value=>{store=value},
+  live:async()=>({image:['old.safetensors','klein-pokemon_anime_style_klein9b.safetensors'],video:['old-vid.safetensors']}),
+  fresh:false
+ })
+ assert.ok(next.image.includes('klein-pokemon_anime_style_klein9b.safetensors'))
+ assert.deepEqual(store.image,next.image)
+ const removed=await cachedLoras({
+  read:()=>store,
+  write:value=>{store=value},
+  live:async()=>({image:['klein-pokemon_anime_style_klein9b.safetensors'],video:[]}),
+  fresh:true
+ })
+ assert.deepEqual(removed,{image:['klein-pokemon_anime_style_klein9b.safetensors'],video:[]})
+})
+test('bench mounts and Refresh LoRAs request a live scan',()=>{
+ const src=readFileSync(new URL('../components/studio2/Bench.vue',import.meta.url),'utf8')
+ assert.match(src,/void loadLoras\(true\)/)
+ assert.match(src,/loadLoras\(true\)/)
+ assert.match(src,/\/api\/studio-2\/loras'\+\(fresh\?'\?fresh=1':''\)/)
 })
 test('source/hero actions are independent and full input restores structured settings',()=>{
  const form={identityStillId:'hero',imageAId:'source',promptSections:{action:'old'},settings:{}}
@@ -51,3 +74,10 @@ test('image chips compile positive and negative independently; video chips do no
  assert.doesNotMatch(compilePrompt(q.promptSections,false,q),/Picture 1/)
  assert.deepEqual(parseBatch('A wrapped\nsentence\n\nSecond shot\r\n\r\nThird'),['A wrapped\nsentence','Second shot','Third'])
 })
+test('qwen21 is generate-only T2I with 25/1 defaults and no locks',()=>{
+ const q=validateRequest({mode:'generate',engine:'qwen21',folderId:'f',promptSections:{action:'a red bicycle'},settings:{aspect:'1:1'}})
+ assert.equal(q.engine,'qwen21');assert.equal(q.settings.steps,25);assert.equal(q.settings.cfg,1)
+ assert.equal(q.settings.width,1024);assert.equal(q.settings.height,1024)
+ assert.equal(q.lockFace,false);assert.equal(q.lockOutfit,false);assert.deepEqual(q.settings.loraStack,[])
+ assert.throws(()=>validateRequest({mode:'edit',engine:'qwen21',folderId:'f',imageAId:'s',promptSections:{action:'x'}}),/T2I in this build/)
+})

tests/studio2-controls.test.mjs

@@ -36,6 +36,17 @@ test('appearance roles distinguish wardrobe lock and both-off graph excludes her
  const graph=nativeVideoGraph(template,r,{hero:'ignored.png',start:'start.png'},'preview')
  assert.equal(graph.hero,undefined);assert.equal(graph['155'].inputs['ref_images.ref_image_0'],undefined);assert.equal(graph.load_start.inputs.image,'start.png');assert.equal(r.identityStillId,'hero')
 })
+test('video with locks off and no assets is MiniMax T2V empty first frame',()=>{
+ const r=validateRequest({mode:'video',engine:'minimax',folderId:'f',lockFace:false,lockOutfit:false,identityStillId:'',imageAId:'',promptSections:{action:'Wide empty street at dawn'}})
+ assert.equal(r.lockFace,false);assert.equal(r.lockOutfit,false);assert.equal(r.imageAId,'');assert.equal(r.startClipId,'')
+ const template=JSON.parse(fs.readFileSync(new URL('../server/assets/studio2_minimax_native.json',import.meta.url)))
+ const graph=nativeVideoGraph(template,r,{hero:'',start:''},'preview/t2v')
+ assert.equal(graph['155'].class_type,'MiniMaxH3ImageToVideo')
+ assert.equal(graph['155'].inputs.first_frame,undefined)
+ assert.equal(graph.load_start,undefined)
+ assert.equal(graph.hero,undefined)
+ assert.throws(()=>validateRequest({mode:'extend',engine:'minimax',folderId:'f',lockFace:false,lockOutfit:false,promptSections:{action:'Continue'}}),e=>/Previous clip|family strip/i.test(e.message))
+})
 test('sound/dialogue/music compile without empty invented audio',()=>{
  const empty=compilePrompt({},true);assert.doesNotMatch(empty,/Soundscape:|<d>/);assert.match(empty,/non_diegetic_music: N\/A/)
  const text=compilePrompt({soundscapeChips:['footsteps'],dialogue:{mode:'spoken off-screen',line:'Hello'}},true);assert.match(text,/lips remain closed/);assert.match(text,/<Subject 1> \(S1\) says <d>\[English\]Hello<\/d>/)

tests/studio2-status-playback.test.mjs

@@ -10,6 +10,8 @@ test('waiting records never replace the idle GPU status after completion',()=>{
  assert.deepEqual(gpuStatus([waiting,{...running,state:'complete'}],health),{label:'Ready',tone:'ready',detail:'No job running'});
  assert.equal(gpuStatus([waiting],health,[{status:'running',name:'Upscale current clip'}]).label,'Upscaling');
  assert.equal(gpuStatus([waiting],{comfy:{asleep:true}}).label,'Sleeping');
+ assert.equal(gpuStatus([waiting],{comfy:{asleep:true},yue2:{busy:true}}).label,'Generating');
+ assert.equal(gpuStatus([waiting],{comfy:{asleep:true}},[],[{status:'running',name:'song',live:{message:'Planning'}}]).label,'Generating');
 });
 test('Unicode video titles produce valid response headers and retain UTF-8 filename',()=>{
  for(const title of ['A scene… “hello” — 🎥.mp4','日本語.mp4','bad\r\nheader".mp4']){

tests/test_yue2_worker.py

@@ -1,7 +1,9 @@
 """CPU-only YuE2 worker regression tests; never import torch or run inference."""
 import importlib.util
+import os
 from pathlib import Path
 import unittest
+from types import SimpleNamespace
 
 spec = importlib.util.spec_from_file_location('yue2_worker', Path(__file__).parents[1] / 'scripts/yue2-worker.py')
 worker = importlib.util.module_from_spec(spec)
@@ -9,12 +11,36 @@ spec.loader.exec_module(worker)
 
 
 class WorkerTests(unittest.TestCase):
-    def test_lyrics_preserve_words_and_normalize_ui_headings(self):
-        self.assertEqual(worker.normalize_lyrics('[Pre-Chorus]\nEvery word stays\n[Outro]\n'), '[prechorus]\nEvery word stays\n\n')
-        self.assertEqual(worker.normalize_lyrics('[Verse 1]\nHello'), '[verse1]\nHello\n\n')
-        for lyrics in ['', 'No heading', '[Verse]\nA\n[Chorus]\nB']:
+    def test_lyrics_allow_multiple_sections_and_wrap_plain_text(self):
+        multi = '[Verse 1]\nHello\n\n[Chorus]\nSing it'
+        self.assertEqual(worker.normalize_lyrics(multi), multi + '\n')
+        self.assertEqual(worker.normalize_lyrics('[Verse 1]\nHello'), '[Verse 1]\nHello\n')
+        self.assertEqual(worker.normalize_lyrics('No heading yet'), '[song]\nNo heading yet\n\n')
         with self.assertRaises(ValueError):
-                worker.normalize_lyrics(lyrics)
+            worker.normalize_lyrics('')
+        with self.assertRaises(ValueError):
+            worker.normalize_lyrics('   ')
+
+    def test_attention_backend_never_selects_flash(self):
+        torch_mod = SimpleNamespace(
+            cuda=SimpleNamespace(is_available=lambda: True),
+            backends=SimpleNamespace(cudnn=SimpleNamespace(is_available=lambda: True)),
+        )
+        attention, pipeline = worker.resolve_attention_backend(torch_mod)
+        self.assertEqual(attention, 'sdpa')
+        self.assertEqual(pipeline, 'torch-eager')
+        self.assertNotEqual(attention, 'flash')
+        self.assertTrue(os.environ.get('PYTORCH_CUDA_ALLOC_CONF', '').startswith('expandable_segments'))
+
+    def test_target_length_maps_to_semantic_max_tokens(self):
+        # 48000 / 1920 = 25 latent frames per second.
+        self.assertEqual(max(200, min(9000, 30 * 25)), 750)
+        self.assertEqual(max(200, min(9000, 15 * 25)), 375)
+        self.assertEqual(max(200, min(9000, 150 * 25)), 3750)
+
+    def test_cuda_oom_detection(self):
+        self.assertTrue(worker.is_cuda_oom(RuntimeError('CUDA out of memory. Tried to allocate 2.49 GiB')))
+        self.assertFalse(worker.is_cuda_oom(RuntimeError('bad lyrics')))
 
 
 if __name__ == '__main__':

tests/yue2.test.mjs

@@ -11,10 +11,15 @@ import { createYueGpHost, validateYueGpRequest } from '../scripts/yuegp-host.mjs
 
 const request = { id: 'test-song-123456', tags: 'pop, warm vocals', lyrics: '[Verse 1]\nA quiet morning, a new day', seed: 0 }
 
-test('YuE2 defaults are 60 seconds and keep Yue lyric validation', () => {
+test('YuE2 defaults are 60 seconds and allow multiple lyric sections', () => {
   assert.deepEqual(validateYue2Request(request), { ...request, duration: 60 })
-  assert.throws(() => validateYue2Request({ ...request, lyrics: '[Verse]\nA\n[Chorus]\nB' }))
-  assert.throws(() => validateYue2Request({ ...request, duration: 20 }))
+  assert.deepEqual(
+    validateYue2Request({ ...request, lyrics: '[Verse]\nA\n[Chorus]\nB' }).lyrics,
+    '[Verse]\nA\n[Chorus]\nB'
+  )
+  assert.throws(() => validateYue2Request({ ...request, lyrics: '' }))
+  assert.throws(() => validateYue2Request({ ...request, duration: 10 }))
+  assert.deepEqual(validateYue2Request({ ...request, duration: 15 }).duration, 15)
   assert.throws(() => validateYue2Request({ ...request, id: '../escape' }))
 })
 
@@ -87,9 +92,10 @@ test('music workflow refuses yue and yue2 while ACE graphs still build', () => {
   assert.equal(existsSync(new URL('../server/assets/workflow_yue.json', import.meta.url)), false)
 })
 
-test('musicChain routes yue2 to startYue2Job and leaves yue on YuEGP', () => {
+test('musicChain is YuE2-only', () => {
   const source = readFileSync(new URL('../server/utils/musicChain.ts', import.meta.url), 'utf8')
-  assert.match(source, /engine === 'yue'\) return startYueGpJob/)
-  assert.match(source, /engine === 'yue2'\) return startYue2Job/)
+  assert.match(source, /startYue2Job/)
   assert.match(source, /from '\.\/yue2'/)
+  assert.doesNotMatch(source, /startYueGpJob/)
+  assert.doesNotMatch(source, /buildMusicWorkflow/)
 })

utils/generationPresets.ts

@@ -5,14 +5,10 @@ import type { ImageV2PresetSettings } from '~/utils/imageV2'
 export type GenerationPresetKind = 'video' | 'image' | 'imagev2' | 'music'
 
 export type MusicPresetSettings = {
-  yueProfile?: 1 | 3
   tags?: string
   lyrics?: string
-  instrumental?: boolean
   duration?: number
-  steps?: number
   seed?: number
-  cfg?: number
   engine?: import('~/utils/music').MusicEngine
 }
 

utils/libraryGeneration.ts

@@ -298,10 +298,20 @@ export function trackSettingsRows(track: {
   lyricsStrength?: number
 }) {
   const rows: SettingsRow[] = []
+  const engine = String(track.engine || '')
+  const isYue2 = engine === 'yue2' || engine === 'yue-2'
   if (track.engine) {
     rows.push({
       label: 'Engine',
-      value: track.engine === 'ace-step-1.5' ? 'ACE-Step 1.5' : track.engine === 'yue' ? 'YuE' : 'ACE-Step'
+      value: isYue2
+        ? 'YuE2'
+        : engine === 'ace-step-1.5'
+          ? 'ACE-Step 1.5 (legacy)'
+          : engine === 'yue'
+            ? 'YuE (legacy)'
+            : engine.startsWith('ace')
+              ? 'ACE-Step (legacy)'
+              : engine
     })
   }
   if (track.tags?.trim()) rows.push({ label: 'Tags', value: track.tags.trim() })
@@ -309,14 +319,17 @@ export function trackSettingsRows(track: {
   else if (track.lyrics?.trim()) rows.push({ label: 'Lyrics', value: track.lyrics.trim() })
   if (typeof track.duration === 'number' && track.duration > 0) {
     const rounded = Math.round(track.duration * 10) / 10
-    rows.push({ label: 'Duration', value: Number.isInteger(rounded) ? `${rounded}s` : `${rounded.toFixed(1)}s` })
+    rows.push({
+      label: isYue2 ? 'Length' : 'Duration',
+      value: Number.isInteger(rounded) ? `${rounded}s` : `${rounded.toFixed(1)}s`
+    })
   }
   if (track.yueProfile != null) rows.push({ label: 'YuEGP profile', value: String(track.yueProfile) })
-  if (track.yueProfile == null && track.steps != null) rows.push({ label: 'Steps', value: String(track.steps) })
-  if (track.yueProfile == null && track.cfg != null) rows.push({ label: 'CFG', value: formatNum(track.cfg) })
-  if (track.samplerName) rows.push({ label: 'Sampler', value: track.samplerName })
-  if (track.scheduler) rows.push({ label: 'Scheduler', value: track.scheduler })
-  if (track.lyricsStrength != null) rows.push({ label: 'Lyric strength', value: formatNum(track.lyricsStrength) })
+  if (!isYue2 && track.yueProfile == null && track.steps != null) rows.push({ label: 'Steps', value: String(track.steps) })
+  if (!isYue2 && track.yueProfile == null && track.cfg != null) rows.push({ label: 'CFG', value: formatNum(track.cfg) })
+  if (!isYue2 && track.samplerName) rows.push({ label: 'Sampler', value: track.samplerName })
+  if (!isYue2 && track.scheduler) rows.push({ label: 'Scheduler', value: track.scheduler })
+  if (!isYue2 && track.lyricsStrength != null) rows.push({ label: 'Lyric strength', value: formatNum(track.lyricsStrength) })
   if (track.seed != null) rows.push({ label: 'Seed', value: String(track.seed) })
   return rows
 }

utils/loras.ts

@@ -283,7 +283,7 @@ export function kleinLoraOnKreaMessage(name: string) {
   return imageLoraEngineMismatchMessage(name, 'krea')
 }
 
-export type LoraUse = 'minimax' | 'ltx' | 'flux' | 'krea'
+export type LoraUse = 'minimax' | 'ltx' | 'flux' | 'krea' | 'qwen21'
 
 export function isSystemGraphLora(name: string) {
   const id = loraIdentityKey(name)
@@ -301,16 +301,19 @@ export function inferLoraUse(name: string): LoraUse[] | null {
   if (/ltx/.test(value)) uses.push('ltx')
   if (/klein|flux/.test(value)) uses.push('flux')
   if (/krea/.test(value)) uses.push('krea')
+  if (/qwen/.test(value)) uses.push('qwen21')
   return uses.length ? uses : null
 }
 
 export function loraMatchesUse(name: string, use: LoraUse) {
+  if (use === 'qwen21') return false
   return !isSystemGraphLora(name) && !!inferLoraUse(name)?.includes(use)
 }
 
 /** Saved stacks may contain unlabelled adapters; never carry a known wrong engine. */
 export function filterLoraStackForUse(stack: unknown, use: LoraUse, xaigen: boolean) {
   if (use === 'ltx' && !xaigen) return []
+  if (use === 'qwen21') return []
   return filterLoraStackForStudio(stack, xaigen).filter(item => {
     const uses = inferLoraUse(item.name)
     return !isSystemGraphLora(item.name) && (!uses || uses.includes(use))
@@ -318,6 +321,7 @@ export function filterLoraStackForUse(stack: unknown, use: LoraUse, xaigen: bool
 }
 
 export function filterLorasForUse(names: string[], use: LoraUse) {
+  if (use === 'qwen21') return []
   return names.filter(name => loraMatchesUse(name, use))
 }
 

utils/music.ts

@@ -1,6 +1,13 @@
-export const MUSIC_DURATION_MIN = 30
-export const MUSIC_DURATION_MAX = 180
+export const MUSIC_DURATION_MIN = 15
+export const MUSIC_DURATION_MAX = 150
 export const MUSIC_DURATION_DEFAULT = 60
+/** YuE2 VAE downsampling: 48000 / 1920 = 25 semantic tokens per second of audio. */
+export const YUE2_TOKENS_PER_SECOND = 25
+
+export function yue2MaxTokens(durationSec: number) {
+  const seconds = clampMusicDuration(durationSec)
+  return Math.max(200, Math.min(9000, seconds * YUE2_TOKENS_PER_SECOND))
+}
 export const MUSIC_STEPS_MIN = 8
 export const MUSIC_STEPS_MAX = 100
 export const MUSIC_STEPS_DEFAULT = 50
@@ -33,27 +40,18 @@ export const DEFAULT_MUSIC_LYRICS = `[Verse 1]
 [Outro]
 `
 
-export type MusicEngine = 'ace-step' | 'ace-step-1.5' | 'yue' | 'yue2'
+export type MusicEngine = 'yue2'
 
 export function parseMusicEngine(raw: unknown): MusicEngine {
   const value = String(raw || '').trim().toLowerCase()
-  if (value === 'ace-step-1.5' || value === 'ace15' || value === 'v1.5') return 'ace-step-1.5'
-  if (value === 'yue2' || value === 'yue-2') return 'yue2'
-  if (value === 'yue' || value === 'yue-s1' || value === 'lyrics2song') return 'yue'
-  return 'ace-step'
-}
-
-export function isAceMusicEngine(engine: MusicEngine | string | undefined) {
-  const value = parseMusicEngine(engine)
-  return value !== 'yue' && value !== 'yue2'
+  if (value === 'yue2' || value === 'yue-2' || !value) return 'yue2'
+  // Legacy ACE / YuEGP values map to YuE2 — the only supported music engine.
+  return 'yue2'
 }
 
 export function musicEngineLabel(engine: MusicEngine | string | undefined) {
-  const value = parseMusicEngine(engine)
-  if (value === 'ace-step-1.5') return 'ACE-Step 1.5'
-  if (value === 'yue2') return 'YuE2'
-  if (value === 'yue') return 'YuE'
-  return 'ACE-Step'
+  void engine
+  return 'YuE2'
 }
 
 export function parseMusicBpm(tags: unknown, fallback = 120) {
@@ -97,18 +95,17 @@ export function sanitizeMusicLyrics(raw: unknown, instrumental = false) {
   return String(raw || '').replace(/\r\n/g, '\n').trim().slice(0, MUSIC_LYRICS_MAX)
 }
 
-/** YuE splits on [Section] headers. Returns how many sections are present. */
+/** YuE2 splits on [Section] headers. Returns how many section tags are present. */
 export function countYueLyricSections(lyrics: unknown) {
   const text = String(lyrics || '')
   const matches = text.match(/\[[^\]]+\]/g)
   return matches ? matches.length : 0
 }
 
+/** YuE2 allows multiple lyric sections; only empty lyrics are rejected. */
 export function yueLyricsProblem(lyrics: unknown): string | null {
   const text = String(lyrics || '').trim()
-  if (!text) return 'YuE needs lyrics with section tags like [Verse 1] and [Chorus].'
-  const sections = [...text.matchAll(/\[([^\]]+)\]\s*([^\[]*)/gs)].filter(match => match[2].trim())
-  if (sections.length !== 1) return 'YuEGP needs exactly one non-empty lyric section. Combine the lyrics under one heading.'
+  if (!text) return 'YuE2 needs lyrics. Use section tags like [Verse 1] and [Chorus] when you want structure.'
   return null
 }