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>

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' : '',
-          dragOverIndex === index ? 'border-amber-300/50' : 'border-white/10',
           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 items-start justify-between gap-2">
-          <div class="flex min-w-0 items-center 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, $event)"
                 />
+              </span>
+              <span v-if="filenameSubtitle" class="mt-0.5 block truncate text-[11px] text-[#9A9588]" :title="item.name">{{ stemFor(item.name) }}</span>
+            </span>
           </div>
           <button
             type="button"
-            class="shrink-0 text-[11px] text-zinc-500 hover:text-red-300 disabled:opacity-50"
+            class="shrink-0 text-[11px] text-[#9A9588] hover:text-red-300 disabled:opacity-50"
             :disabled="disabled"
             @click="removeAt(index)"
           >
             Remove
           </button>
         </div>
-        <div class="mt-1.5 flex flex-wrap gap-3">
+
-          <Thumbwheel
+        <div class="mt-2 space-y-1.5">
-            :model-value="item.strengthModel"
+          <label class="grid grid-cols-[3.25rem_minmax(0,1fr)_2.75rem] items-center gap-2 text-[11px] text-[#9A9588]">
-            label="Model"
+            <span>{{ clipStrength ? 'Model' : 'Strength' }}</span>
-            :lock-key="`lora-model:${item.name}`"
+            <input
-            :min="LORA_STRENGTH_MIN"
+              type="range"
-            :max="LORA_STRENGTH_MAX"
+              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)"
+              :value="displayStrength(item.strengthModel)"
-          />
+              :aria-label="(clipStrength ? 'Model' : 'Strength') + ' for ' + labelFor(item.name)"
-          <Thumbwheel
+              @input="onRange(index, 'strengthModel', $event)"
-            :model-value="item.strengthClip"
+            >
-            label="CLIP"
+            <input
-            :lock-key="`lora-clip:${item.name}`"
+              type="number"
-            :min="LORA_STRENGTH_MIN"
+              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]"
-            :max="LORA_STRENGTH_MAX"
+              :min="0"
+              :max="2"
               :step="LORA_STRENGTH_STEP"
-            :fine-step="LORA_STRENGTH_STEP"
+              :disabled="disabled"
-            :coarse-step="0.25"
+              :value="displayStrength(item.strengthModel).toFixed(2)"
-            :digits="2"
+              :aria-label="(clipStrength ? 'Model' : 'Strength') + ' value'"
-            compact
+              @change="onNumber(index, 'strengthModel', $event)"
-            :disabled="disabled || !clipStrength"
+            >
-            @update:model-value="setStrength(index, 'strengthClip', $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_DEFAULT,
-  LORA_STRENGTH_MIN,
   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)
+  const next = clampUi(value)
-  model.value = model.value.map((item, i) => i === index ? { ...item, [key]: next } : item)
+  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/studio2/Bench.vue

@@ -11,7 +11,7 @@
       <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'" class="flex flex-wrap items-center gap-3 text-sm text-zinc-300" aria-label="Identity locks">
+    <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>
@@ -45,7 +45,7 @@
     >{{ task }}</button>
    </nav>
 
-   <div v-if="form.mode!=='music'" class="min-w-0">
+   <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>
@@ -68,7 +68,8 @@
      </button>
     </div>
    </div>
-   <p v-else class="display-body text-zinc-400">Model · YuE2</p>
+   <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">
@@ -103,6 +104,7 @@
      <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" />
@@ -139,12 +141,14 @@
      </div>
     </div>
 
-    <section v-if="form.mode!=='music'" aria-label="Source" class="display-card">
+    <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">
@@ -192,6 +196,18 @@
       <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" />
@@ -206,6 +222,7 @@
      </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>
+     <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>
 
@@ -220,6 +237,25 @@
      <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>
@@ -230,7 +266,7 @@
      <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)) || musicJobs.some(j=>['complete','done','error','cancelled'].includes(j.status))" type="button" class="text-sm text-zinc-400 outline-none ring-amber-300/40 focus-visible:ring-2" @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>
 
@@ -249,8 +285,10 @@
      </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>
-     <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 += ' '+$event" />
+     <template v-if="form.engine!=='qwen21'">
+     <LoraSelect v-model="form.settings.loraStack" class="mt-5" :names="pickerLoras" :use="currentLoraUse" :clip-strength="!video" filename-subtitle :aliases="{...loraNames.aliases,...library.loraAliases}" :triggers="{...loraNames.triggers,...library.loraTriggers}" missing-label="No cached LoRAs yet — Refresh to check Comfy" @insert-trigger="form.promptSections.action += ' '+$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>
@@ -283,9 +321,10 @@
 import {createLoraMemory,loraUse} from '~/shared/studio2/lora-memory.mjs'
 import {filterLorasForStudio,filterLorasForUse} from '~/utils/loras'
 import {compilePrompt,sectionsFrom} from '~/shared/studio2/contracts.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} from '~/shared/studio2/queue-view.mjs'
+import {queueJobs,elapsedSeconds} from '~/shared/studio2/queue-view.mjs'
 import {gpuStatus} from '~/shared/studio2/status.mjs'
 import {applyUploadedStill} from '~/shared/studio2/upload.mjs'
 import {parseBatch,stylePrompt} from '~/shared/studio2/styles.mjs'
@@ -307,15 +346,19 @@ 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','music']
+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 [
@@ -325,30 +368,78 @@ const engineOptions=computed(()=>{
  }
  return [
   {value:'flux',label:'Klein',disabled:false,hint:''},
-  {value:'krea',label:'Krea 2',disabled:kreaSticky.value,hint:kreaSticky.value?'sticky hero = Klein':''}
+  {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/${id}`
 const folderJobs=computed(()=>jobs.value.filter(j=>j.request.folderId===form.folderId))
-const benchJobs=computed(()=>queueJobs(folderJobs.value))
+const captionJobs=computed(()=>folderJobs.value.filter(j=>j.request?.mode==='caption'))
+const benchJobs=computed(()=>queueJobs(folderJobs.value.filter(j=>j.request?.mode!=='caption')))
 const musicJobs=computed(()=>studioQueue.value
  .filter((j:any)=>j.kind==='music' && j.folderId===form.folderId && !['complete','done'].includes(j.status))
  .map((j:any)=>({...j,live:j.liveJobId?liveMusic.value[j.liveJobId]:null})))
-const gpu=computed(()=>gpuStatus(jobs.value,healthState.value,upscales.value,musicJobs.value))
+const gpu=computed(()=>gpuStatus(jobs.value,healthState.value,upscales.value,musicJobs.value,captionJobs.value))
+const captionStill=computed(()=>stills.value.find((s:any)=>s.id===captionStillId.value))
 const sourceSlot=computed(()=>{const clip=clips.value.find((a:any)=>a.id===form.startClipId);return form.mode==='extend'?(clip?{...clip,kind:'clip'}:null):inputStill.value})
 const pickerAssets=computed(()=>picker.value==='startClipId'?clips.value.map((a:any)=>({...a,kind:'clip'})):stills.value.map((a:any)=>({...a,kind:'still'})))
 const inputStill=computed(()=>stills.value.find((s:any)=>s.id===form.imageAId))
 const family=computed(()=>folderAssets(library.value,jobs.value,form.folderId))
 const sourceAsset=computed(()=>form.startClipId?clips.value.find((c:any)=>c.id===form.startClipId):stills.value.find((s:any)=>s.id===(form.imageAId||form.identityStillId)))
 const sourceSize=computed(()=>measured[sourceAsset.value?.id] || sourceAsset.value)
-const resolvedSize=computed(()=>resolveSize(form.settings.aspect,sourceSize.value))
+const resolvedSize=computed(()=>resolveSize(form.settings.aspect,sourceSize.value,undefined,form.engine))
+
+function elapsedCaption(job:any){
+ const seconds=elapsedSeconds(job,now.value)
+ return `${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
@@ -380,6 +471,11 @@ watch(()=>sourceAsset.value?.id,id=>{const asset=sourceAsset.value;if(!import.me
 
 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'
@@ -398,7 +494,7 @@ function selectOutput(item:any){selected.value=item;if(form.mode==='extend' && i
 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
@@ -416,6 +512,10 @@ async function queue(){
    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
@@ -470,6 +570,7 @@ onMounted(async()=>{
  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)})

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`.

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

@@ -4,6 +4,7 @@ import { stableMemoryArgs } from './comfy-memory-policy.mjs'
 import { createGpuReservation } from './gpu-reservation.mjs'
 import { createGpuProxy } from './gpu-proxy.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'
@@ -147,7 +148,7 @@ let proxyTarget = 0
 
 function ensureProxyListening() {
   if (proxyServer) return
-  proxyServer = createGpuProxy({ target: () => proxyTarget, reservation: gpuReservation, authorized, markWork, externalBusy: () => 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) }))
   })
@@ -819,7 +820,7 @@ function purgeDesktopFiles(body) {
 }
 
 const gpuReservation = createGpuReservation({ idle: async () => {
-  if (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) {
@@ -830,14 +831,15 @@ const gpuReservation = createGpuReservation({ idle: async () => {
   return !(await processUp()) && !(await pythonMainUp().catch(() => true))
 } })
 
-const yue2 = createYue2Host({
+async function prepareExclusiveGpu(label) {
-  leaseValid: lease => gpuReservation.isOwner(lease),
-  prepare: async () => {
   if ((await trainingLock()).busy) throw new Error('GPU is busy with training.')
+  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; YuE2 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)) {
@@ -846,13 +848,50 @@ const yue2 = createYue2Host({
   }
   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 () => prepareExclusiveGpu('YuE2')
+})
+
+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'] || '')))
@@ -909,12 +948,14 @@ async function handleControl(req, res) {
       gpu: gpuReservation.availability(),
       training: { busy: lastTraining.busy },
       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 (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, {
@@ -1016,7 +1057,7 @@ const server = http.createServer(async (req, res) => {
     } else await handleControl(req, res)
   } catch (error) {
     req.resume()
-    if ((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/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/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())

server/api/health.get.ts

@@ -48,7 +48,7 @@ export default defineEventHandler(async () => {
       host: comfy.host || '',
       queueRunning: queue.running,
       queuePending: queue.pending,
-      asleep: Boolean(control?.asleep) && !comfyReachable && !control?.yue2?.busy,
+      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
@@ -58,6 +58,11 @@ export default defineEventHandler(async () => {
       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/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/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/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/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';
@@ -126,6 +128,28 @@ async function prepareGraph(r: any) {
             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'
 }
 
@@ -288,7 +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')
@@ -308,12 +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')
@@ -353,7 +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
@@ -373,7 +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') {
@@ -467,8 +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 kind = params.kind === 'edit' ? 'edit'
-  const shotCount = kind === 'music'
+    : params.kind === 'music' ? 'music'
+      : params.kind === 'caption' ? 'caption'
+        : 'video'
+  const shotCount = kind === 'music' || kind === 'caption'
     ? 1
     : kind === 'edit'
       ? 1 + (params.payload.passes?.length || 0)
@@ -604,6 +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) {
@@ -676,6 +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) {
@@ -887,6 +901,10 @@ function pendingAlive(job: StudioJob) {
       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 => (
@@ -1395,6 +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 {
@@ -1437,6 +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')
@@ -1701,7 +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) => {

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.')
@@ -46,8 +47,13 @@ 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}.`)
+  const qwen = engine === 'qwen21'
-  for (const key of ['width','height']) if (number(key,key==='width'?1344:768,256,4096)%32) fail('Dimensions must be multiples of 32.')
+  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.')
@@ -55,10 +61,10 @@ export function validateRequest(raw, xaigen = false) {
   }
   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),
+    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',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:[] },
+    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/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-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

@@ -1,4 +1,4 @@
-export function gpuStatus(jobs, health, upscales=[], musicJobs=[]) {
+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'}
@@ -7,6 +7,9 @@ export function gpuStatus(jobs, health, upscales=[], musicJobs=[]) {
   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'}

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

@@ -74,3 +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/)
+})

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))
 }