Keep last-frame continuity when character stills are attached.

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

docs/image-v2.md

@@ -100,7 +100,7 @@ Generate Krea (`krea_v2_generate.json`):
 
 Krea Edit / Compose / Refine use the same UNET / CLIP / VAE. They encode the still (`VAEEncode`) and sample with denoise < 1. Krea has no Klein `ReferenceLatent`. Compose stitches A then B (`ImageStitch`) so both stills are in the latent. Missing Krea models fail the job. No Flux fallback.
 
-Krea does not load Klein UNET, Klein CLIP, Klein VAE, Klein Consistency, or the Klein concept LoRA. Concept LoRA (`klein_snofs_v1_4`) is xAIGen only; AIGen bypasses that node. On Krea it stays off unless `NUXT_KREA_CONCEPT_LORA` is set on xAIGen.
+Krea does not load Klein UNET, Klein CLIP, Klein VAE, Klein Consistency, or the Klein concept LoRA. Concept LoRA (`xaigen-klein_snofs_v1_4`) is xAIGen only; AIGen bypasses that node. On Krea it stays off unless `NUXT_KREA_CONCEPT_LORA` is set on xAIGen.
 
 There is no `LoadImage` on Generate. If the executed Generate graph has a required LoadImage, the job fails. v1 node 145 (muted T2I with a baked prompt) is not used.
 

pages/index.vue

@@ -273,10 +273,10 @@
           />
           <p v-if="shotScriptMode && studioMode === 'video'" class="mb-2 text-xs text-zinc-500">
             Paste a block, or load a saved Director script. Text before the first marker is shot 1. Put <span class="font-medium text-zinc-300">shot 2</span>, <span class="font-medium text-zinc-300">shot 3</span> on their own lines to queue extensions — a line like <span class="font-medium text-zinc-300">Shot Type</span> will not start a new shot. Every shot uses the duration slider. Global locks are inherited; each shot should add one beat.
-            <span v-if="identityPrompting"> The Comfy identity prefix is prepended to every shot, and the same Picture 1 / extra stills are sent with each shot.</span>
+            <span v-if="identityPrompting"> Character stills ride along as extra pictures. Picture 1 is the current frame (last frame after shot 1), not the original start still.</span>
           </p>
           <p v-else-if="identityPrompting" class="mb-2 text-xs text-zinc-500">
-            Write camera, action, and audio. That is the dynamic action prompt. We concatenate the <span class="font-medium text-zinc-300">&lt;Picture 1&gt;</span> identity prefix in front of it and send the result to <span class="font-medium text-zinc-300">Reference to Video</span>, not Image to Video.
+            Write camera, action, and audio. Picture 1 is the current frame. Extra stills lock face and body only — they do not reset clothes. MiniMax still cannot hard-lock a last frame and character refs in one node, so this path is Reference to Video with the last frame as Picture 1.
           </p>
           <div
             class="grid gap-3"
@@ -667,25 +667,18 @@
                 :class="minimaxGraph === 'v2' ? 'border-amber-300 bg-amber-400/10 text-amber-100' : 'border-white/10 hover:border-white/20'"
                 @click="minimaxGraph = 'v2'"
               >
-                <span class="block font-semibold">v2 · identity refs</span>
+                <span class="block font-semibold">v2 · last frame + stills</span>
-                <span class="text-xs text-zinc-400">Optional lock with up to 4 stills</span>
+                <span class="text-xs text-zinc-400">Optional character stills. Last frame stays the start.</span>
               </button>
             </div>
           </div>
 
           <div v-if="videoEngine === 'minimax' && videoStart === 'still' && minimaxGraph === 'v2'" class="rounded-2xl border border-white/10 bg-zinc-950/40 px-4 py-3 space-y-3">
-            <label class="flex cursor-pointer items-center justify-between gap-3 text-sm" :class="extraPersonLock ? 'cursor-not-allowed opacity-60' : ''">
+            <div>
-              <span>
+              <p class="text-sm font-medium text-zinc-200">Character stills</p>
-                <span class="block font-medium text-zinc-200">Use identity references</span>
+              <p class="mt-0.5 text-xs text-zinc-500">Optional. Extra views of the same person — face, hair, body. Shot 2+ still starts from the last frame (clothes and pose stay). Empty slots are left out. A named extra person turns these off.</p>
-                <span class="mt-0.5 block text-xs text-zinc-500">Off uses last-frame I2V after shot 1. On keeps the start still as Picture 1 — extra stills are more views of that same person, never a second character. A named extra person forces last-frame I2V.</span>
+            </div>
-              </span>
+            <div v-if="!extraPersonLock" class="grid grid-cols-2 gap-2 sm:grid-cols-4">
-              <span class="relative inline-flex h-6 w-11 shrink-0 items-center">
-                <input v-model="useIdentityRefs" type="checkbox" class="peer sr-only" role="switch" :disabled="extraPersonLock">
-                <span class="absolute inset-0 rounded-full bg-zinc-700 transition peer-checked:bg-amber-400" />
-                <span class="absolute left-0.5 top-0.5 h-5 w-5 rounded-full bg-white transition peer-checked:translate-x-5" />
-              </span>
-            </label>
-            <div v-if="useIdentityRefs && !extraPersonLock" class="grid grid-cols-2 gap-2 sm:grid-cols-4">
               <div
                 v-for="slot in 4"
                 :key="slot"
@@ -723,8 +716,8 @@
                 </button>
               </div>
             </div>
-            <p v-if="extraPersonLock" class="text-xs text-zinc-500">Identity refs stay off while a second named person is attached. MiniMax will not see the extra stills; shot 2+ starts from the parent clip’s last frame.</p>
+            <p v-if="extraPersonLock" class="text-xs text-zinc-500">Character stills stay off while a second named person is attached. Shot 2+ starts from the last frame.</p>
-            <p v-else-if="useIdentityRefs" class="text-xs text-zinc-500">Picture 1 is the start still and stays Picture 1 on every shot, including queued extensions. Extra stills map to Picture 2–5 only if you add them — extra views of that same person, not a second character or a vehicle. Empty slots are left out of the graph, not filled with copies of Picture 1.</p>
+            <p v-else class="text-xs text-zinc-500">Picture 1 is the start still on shot 1 and the last frame after that. Extra stills are Picture 2–5. They do not put the start-still outfit back on.</p>
           </div>
         </div>
 
@@ -2449,7 +2442,6 @@ const imageV2LoraOptions = computed(() => mergeLoraOptions(
   filterLorasForImageEngine(imageLoras.value, v2Engine.value),
   filterLorasForImageEngine(imageV2LoraStack.value.map(item => item.name), v2Engine.value)
 ))
-const useIdentityRefs = ref(false)
 const identityRefs = ref<(File | null)[]>([null, null, null, null])
 const identityPreviews = ref(['', '', '', ''])
 const globalLocks = ref('')
@@ -2474,12 +2466,13 @@ const extraPersonLock = computed(() => shouldForceLastFrameI2V(
   mergePermanenceRefs(familyPermanenceRefs.value, Object.values(shotPermanenceRefs.value).flat()),
   globalLocks.value
 ))
+const useIdentityRefs = computed(() => !extraPersonLock.value && identityRefs.value.some(Boolean))
 const permanenceComfyNote = computed(() => {
   if (extraPersonLock.value) {
     return 'A second named person keeps last-frame I2V on. MiniMax never sees the man/bench photos; those names only go in the inherit line.'
   }
   if (studioMode.value === 'video' && videoWorkflow.value === 'v2' && useIdentityRefs.value) {
-    return 'Identity-ref still only maps extra views of the start-still person, never a second character. Permanence stills name the lock text; they are not Picture 2–5.'
+    return 'Character stills are extra views of the same person. Picture 1 is the last frame after shot 1. Permanence stills only name the lock text.'
   }
   return 'Last-frame I2V does not send these stills to MiniMax as extra pictures. Continuity is the previous clip’s last frame; the names only go in the lock line.'
 })
@@ -2667,7 +2660,7 @@ const videoModelHint = computed(() => {
       : 'PinkCherry image-to-video. Distilled LoRA stays on. Silent MP4 — LTX has no audio VAE here.'
   }
   if (textToVideo.value) return 'MiniMax H3 text-to-video with native stereo audio. Shot 2+ switches to last-frame image-to-video.'
-  if (videoWorkflow.value === 'v2') return 'MiniMax H3 image-to-video. Identity refs stay on v2 with a start still.'
+  if (videoWorkflow.value === 'v2') return 'MiniMax H3 image-to-video. Shot 2+ starts from the last frame. Character stills are optional extras.'
   return 'MiniMax H3 image-to-video with native stereo audio.'
 })
 const frameCount = computed(() => Math.max(5, Math.floor(duration.value * fps.value)))
@@ -3320,7 +3313,6 @@ onMounted(async () => {
         applyEngineDefaults('ltx')
       }
     } catch { /* ignore */ }
-    useIdentityRefs.value = localStorage.getItem('aigen-use-identity-refs') === 'true'
     videoLoraStack.value = readStoredLoraStack('aigen-video-lora')
     imageLoraStack.value = readStoredLoraStack('aigen-image-lora')
     imageV2LoraStack.value = readStoredLoraStack('aigen-image-v2-lora')
@@ -3469,16 +3461,6 @@ watch(videoStart, (value) => {
   } catch { /* ignore */ }
 })
 
-watch(useIdentityRefs, (value) => {
-  try {
-    localStorage.setItem('aigen-use-identity-refs', String(value))
-  } catch { /* ignore */ }
-})
-
-watch(extraPersonLock, (locked) => {
-  if (locked) useIdentityRefs.value = false
-})
-
 watch([globalLocks, familyPermanenceRefs, shotPermanenceRefs], () => {
   try {
     localStorage.setItem(LOCKS_STORE, globalLocks.value)
@@ -4493,7 +4475,6 @@ async function restoreStudioJob(id: string) {
   }
   if (typeof payload.sound === 'boolean') withSound.value = payload.sound
   shotScriptMode.value = extensions.length > 0
-  useIdentityRefs.value = payload.useIdentityRefs === true
   identityRefs.value.forEach((_, index) => clearIdentityRef(index))
   extensionQueue.value = extensions.map(item => ({
     id: crypto.randomUUID(),
@@ -4749,7 +4730,6 @@ async function rerun(item: LibraryClip, collection = false) {
   applyVideoWorkflow(target.workflow)
   if (typeof target.sound === 'boolean') withSound.value = target.sound
   shotScriptMode.value = restoreAll
-  useIdentityRefs.value = false
   identityRefs.value.forEach((_, index) => clearIdentityRef(index))
   const ordered = chronologicalParts(parts)
   globalLocks.value = ordered.find(part => part.globalLocks)?.globalLocks || target.globalLocks || ''

server/assets/klein_v2_compose.json

@@ -58,7 +58,7 @@
   },
   "7": {
     "inputs": {
-      "lora_name": "klein_snofs_v1_4.safetensors",
+      "lora_name": "xaigen-klein_snofs_v1_4.safetensors",
       "strength_model": 0.65,
       "strength_clip": 0.35,
       "model": ["4", 0],

server/assets/klein_v2_edit.json

@@ -43,7 +43,7 @@
   },
   "7": {
     "inputs": {
-      "lora_name": "klein_snofs_v1_4.safetensors",
+      "lora_name": "xaigen-klein_snofs_v1_4.safetensors",
       "strength_model": 0.65,
       "strength_clip": 0.35,
       "model": ["4", 0],

server/assets/klein_v2_generate.json

@@ -23,7 +23,7 @@
   },
   "7": {
     "inputs": {
-      "lora_name": "klein_snofs_v1_4.safetensors",
+      "lora_name": "xaigen-klein_snofs_v1_4.safetensors",
       "strength_model": 0.65,
       "strength_clip": 0.35,
       "model": ["4", 0],

server/assets/klein_v2_refine.json

@@ -76,7 +76,7 @@
   },
   "7": {
     "inputs": {
-      "lora_name": "klein_snofs_v1_4.safetensors",
+      "lora_name": "xaigen-klein_snofs_v1_4.safetensors",
       "strength_model": 0.65,
       "strength_clip": 0.3,
       "model": ["4", 0],

server/assets/workflow_minimax_video_v2.json

@@ -97,7 +97,7 @@
   },
   "155": {
     "inputs": {
-      "prompt": "<Picture 1> is the exact facial identity, hairstyle, body proportions and clothing of the main character. Preserve this identity with high fidelity even when the subject temporarily leaves the frame and returns.\n\n[describe the action, camera and audio here]",
+      "prompt": "<Picture 1> is the current frame: keep this pose, clothes, place, and who is in shot. Extra stills are face, hair, and body only — not a wardrobe change.\n\n[describe the action, camera and audio here]",
       "width": [
         "127",
         0

server/utils/videoChain.ts

@@ -152,9 +152,9 @@ export async function queueMiniMax(
   const hasImage = Boolean(params.image?.data?.length)
   const uploading = !hasImage
     ? `Queueing ${engineName} text-to-video…`
-    : params.useIdentityRefs
+    : chainIndex > 0
-      ? (chainIndex > 0 ? 'Uploading identity stills for next shot...' : 'Uploading image to ComfyUI...')
+      ? (params.useIdentityRefs ? 'Uploading last frame and character stills...' : 'Uploading last frame to ComfyUI...')
-      : (chainIndex > 0 ? 'Uploading last frame to ComfyUI...' : 'Uploading image to ComfyUI...')
+      : 'Uploading image to ComfyUI...'
   const queueing = chainIndex > 0
     ? `Queueing extension on ${engineName}...`
     : `Queueing ${engineName} job...`
@@ -393,9 +393,7 @@ export async function continueQueuedExtensions(
       await ensureComfyReady(ready)
       await queueMiniMax(job, {
         prompt: ext.prompt,
-        image: identity
+        image: { filename: 'last_frame.png', data: frame, type: 'image/png' },
-          ? params.image
-          : { filename: 'last_frame.png', data: frame, type: 'image/png' },
         width: params.width,
         height: params.height,
         steps: params.steps,

utils/identityPrompt.ts

@@ -11,13 +11,13 @@ export function identityBoilerplate(extraRefs: number | number[] = 0) {
   const secondary = extras.length === 0
     ? ''
     : extras.length === 1
-      ? ` <Picture ${extras[0]}> is another view of that same person for identity only, not a second character.`
+      ? ` <Picture ${extras[0]}> is that same person for face, hair, and body only — not a second character and not a wardrobe change.`
-      : ` ${extras.map(n => `<Picture ${n}>`).join(' and ')} are extra views of that same person for identity only, not additional characters.`
+      : ` ${extras.map(n => `<Picture ${n}>`).join(' and ')} are extra views of that same person for face, hair, and body only — not additional characters and not a wardrobe change.`
-  return `<Picture 1> is the identity lock for a single main subject: exact face, hair, body, and clothing.${secondary} Do not spawn extra people from reference stills. If this subject leaves the frame, they must return as the same person from <Picture 1>.`
+  return `<Picture 1> is the current frame: keep this pose, clothes, place, and who is in shot. Do not reset the outfit to an earlier still.${secondary} Do not spawn extra people from reference stills. If this subject leaves the frame, they return as the person from the extra stills, wearing whatever Picture 1 already shows.`
 }
 
 export function looksLikeIdentityPrompt(text: string) {
-  return /^<Picture 1> is the (identity lock|primary identity anchor|exact facial identity)/i.test(String(text || '').trim())
+  return /^<Picture 1> is the (current frame|identity lock|primary identity anchor|exact facial identity)/i.test(String(text || '').trim())
 }
 
 export function buildIdentityPrompt(actionPrompt: string, extraRefs: number | number[] = 0) {

utils/imageV2.ts

@@ -26,7 +26,7 @@ export const IMAGE_V2_STRENGTH_MIN = 0
 export const IMAGE_V2_STRENGTH_MAX = 2
 export const IMAGE_V2_STRENGTH_STEP = 0.05
 
-export const IMAGE_V2_SNOFS_LORA = 'klein_snofs_v1_4.safetensors'
+export const IMAGE_V2_SNOFS_LORA = 'xaigen-klein_snofs_v1_4.safetensors'
 export const IMAGE_V2_CONSISTENCY_LORA = 'Flux2-Klein-9B-consistency-V2.safetensors'
 export const IMAGE_V2_DENOISE_DEFAULT = 0.35
 export const IMAGE_V2_DENOISE_MIN = 0.15