Add MiniMax and PinkCherry LTX as pickable video engines, with still or text start.

Shot 1 can run text-to-video; later shots still extend from the last frame. Recommend output is also normalized into shot-script cards.

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

server/api/generate.post.ts

@@ -1,5 +1,6 @@
 import { frameLength, runGeneration } from '~/server/utils/videoChain'
 import { createShotQueue, finishQueueBurst, setQueueJob } from '~/server/utils/shotQueue'
+import { defaultVideoSteps, isLtxWorkflow, isTextToVideo, parseVideoWorkflow } from '~/utils/videoModels'
 
 function isComfyBusyTimeout(error: unknown) {
   const err = error as { statusCode?: number; status?: number; data?: { code?: string }; message?: string; statusMessage?: string }
@@ -42,16 +43,16 @@ function parseFps(raw: string | undefined) {
   return fps === 12 || fps === 30 || fps === 24 ? fps : 24
 }
 
-function parseCfg(raw: string | undefined, turbo: boolean) {
-  const fallback = turbo ? 1.5 : 4
+function parseCfg(raw: string | undefined, turbo: boolean, ltx = false) {
+  const fallback = ltx ? 3.5 : (turbo ? 1.5 : 4)
   const value = Number(raw)
   if (!Number.isFinite(value)) return fallback
   const snapped = Math.round(value * 2) / 2
   return Math.min(10, Math.max(1, snapped))
 }
 
-function parseSampler(raw: string | undefined) {
-  return SAMPLERS.has(raw || '') ? raw! : 'res_multistep'
+function parseSampler(raw: string | undefined, ltx = false) {
+  return SAMPLERS.has(raw || '') ? raw! : (ltx ? 'euler' : 'res_multistep')
 }
 
 function parseScheduler(raw: string | undefined) {
@@ -86,7 +87,9 @@ export default defineEventHandler(async (event) => {
     }
   }
 
-  if (!image) {
+  const workflow = parseVideoWorkflow(fields.workflow)
+  const textToVideo = isTextToVideo(workflow)
+  if (!image && !textToVideo) {
     throw createError({ statusCode: 400, statusMessage: 'An initial image is required' })
   }
   const prompt = (fields.prompt || '').trim()
@@ -94,10 +97,13 @@ export default defineEventHandler(async (event) => {
     throw createError({ statusCode: 400, statusMessage: 'A motion prompt is required' })
   }
   const extensions = parseExtensions(fields.extensions)
-  const workflow = parseVideoWorkflow(fields.workflow)
-  const useIdentityRefs = workflow === 'v2' && fields.useIdentityRefs === 'true'
-
-  const { width, height } = resolveOutputSize(fields.aspect, image.data, Number(fields.width), Number(fields.height))
+  const useIdentityRefs = workflow === 'v2' && !textToVideo && !isLtxWorkflow(workflow) && fields.useIdentityRefs === 'true'
+  const { width, height } = resolveOutputSize(
+    fields.aspect,
+    image?.data || Buffer.alloc(0),
+    Number(fields.width),
+    Number(fields.height)
+  )
   const turbo = fields.turbo !== 'false'
   const steps = Number(fields.steps || defaultVideoSteps(turbo, workflow))
   const seed = fields.seed && fields.seed !== 'random'
@@ -105,13 +111,13 @@ export default defineEventHandler(async (event) => {
     : Math.floor(Math.random() * 2_147_483_647)
   const durationSeconds = parseDuration(fields.duration || '5')
   const fps = parseFps(fields.fps)
-  const cfg = parseCfg(fields.cfg, turbo)
-  const samplerName = parseSampler(fields.sampler_name)
+  const cfg = parseCfg(fields.cfg, turbo, isLtxWorkflow(workflow))
+  const samplerName = parseSampler(fields.sampler_name, isLtxWorkflow(workflow))
   const scheduler = parseScheduler(fields.scheduler)
   const length = frameLength(durationSeconds, fps)
   const hideThumbnail = fields.hideThumbnail === 'true'
   const hideInput = fields.hideInput === 'true'
-  const sound = fields.sound !== 'false'
+  const sound = !isLtxWorkflow(workflow) && fields.sound !== 'false'
   const clipName = (fields.name || '').trim().slice(0, 80)
   const ownerKey = libraryOwnerKey(event)
   const library = publicLibrary(event)
@@ -124,15 +130,17 @@ export default defineEventHandler(async (event) => {
   const destFolder = library.folders.find(folder => folder.id === folderId)
   const folderLocked = Boolean(destFolder?.protected && !destFolder.unlocked)
   assertFolderExists(event, folderId)
-  const still = await rememberInputStill({
-    ownerKey,
-    folderId,
-    filename: image.filename,
-    data: image.data,
-    width,
-    height,
-    hideInput
-  })
+  const still = image
+    ? await rememberInputStill({
+      ownerKey,
+      folderId,
+      filename: image.filename,
+      data: image.data,
+      width,
+      height,
+      hideInput
+    })
+    : null
   const referenceStillIds: Array<string | null> = [null, null, null, null]
   if (useIdentityRefs) {
     for (const [index, ref] of referenceImages.entries()) {
@@ -175,7 +183,7 @@ export default defineEventHandler(async (event) => {
     fps,
     samplerName,
     scheduler,
-    thumb: isPipelineFrameFilename(image.filename) ? undefined : image.data,
+    thumb: image && !isPipelineFrameFilename(image.filename) ? image.data : undefined,
     stillId: still?.id,
     stillFilename: still?.filename,
     referenceStillIds,

server/assets/workflow_ltx_video.json

@@ -0,0 +1,168 @@
+{
+  "1": {
+    "inputs": {
+      "unet_name": "PinkCherry_FineTune_int8_v1_8_LTX23.safetensors",
+      "weight_dtype": "default"
+    },
+    "class_type": "UNETLoader",
+    "_meta": { "title": "Load PinkCherry LTX-2.3" }
+  },
+  "21": {
+    "inputs": {
+      "lora_name": "ltx-2.3-22b-distilled-lora-384-1.1.safetensors",
+      "strength_model": 1,
+      "model": ["1", 0]
+    },
+    "class_type": "LoraLoaderModelOnly",
+    "_meta": { "title": "LTX distilled LoRA" }
+  },
+  "3": {
+    "inputs": {
+      "vae_name": "LTX23_video_vae_bf16.safetensors"
+    },
+    "class_type": "VAELoader",
+    "_meta": { "title": "Load VAE" }
+  },
+  "4": {
+    "inputs": {
+      "image": "input_image.png",
+      "upload": "image"
+    },
+    "class_type": "LoadImage",
+    "_meta": { "title": "Load Image" }
+  },
+  "5": {
+    "inputs": {
+      "width": 1280,
+      "height": 1280,
+      "interpolation": "lanczos",
+      "method": "keep proportion",
+      "condition": "always",
+      "multiple_of": 32,
+      "image": ["4", 0]
+    },
+    "class_type": "ImageResize+",
+    "_meta": { "title": "Resize still" }
+  },
+  "6": {
+    "inputs": {
+      "text": "",
+      "clip": ["20", 0]
+    },
+    "class_type": "CLIPTextEncode",
+    "_meta": { "title": "Positive prompt" }
+  },
+  "7": {
+    "inputs": {
+      "text": "blurry, low quality, jitter, deformed, static frame, freeze, watermark, bad anatomy, text, subtitles",
+      "clip": ["20", 0]
+    },
+    "class_type": "CLIPTextEncode",
+    "_meta": { "title": "Negative prompt" }
+  },
+  "8": {
+    "inputs": {
+      "width": ["5", 1],
+      "height": ["5", 2],
+      "length": 97,
+      "batch_size": 1,
+      "strength": 1,
+      "positive": ["6", 0],
+      "negative": ["7", 0],
+      "vae": ["3", 0],
+      "image": ["5", 0]
+    },
+    "class_type": "LTXVImgToVideo",
+    "_meta": { "title": "LTX image to video" }
+  },
+  "9": {
+    "inputs": {
+      "frame_rate": 24,
+      "positive": ["8", 0],
+      "negative": ["8", 1]
+    },
+    "class_type": "LTXVConditioning",
+    "_meta": { "title": "LTX conditioning" }
+  },
+  "10": {
+    "inputs": {
+      "cfg": 3.5,
+      "model": ["21", 0],
+      "positive": ["9", 0],
+      "negative": ["9", 1]
+    },
+    "class_type": "CFGGuider",
+    "_meta": { "title": "CFG Guider" }
+  },
+  "11": {
+    "inputs": {
+      "sampler_name": "euler"
+    },
+    "class_type": "KSamplerSelect",
+    "_meta": { "title": "KSamplerSelect" }
+  },
+  "12": {
+    "inputs": {
+      "steps": 8,
+      "max_shift": 2.05,
+      "base_shift": 0.95,
+      "stretch": true,
+      "terminal": 0.1,
+      "latent": ["8", 2]
+    },
+    "class_type": "LTXVScheduler",
+    "_meta": { "title": "LTX scheduler" }
+  },
+  "13": {
+    "inputs": {
+      "noise_seed": 1
+    },
+    "class_type": "RandomNoise",
+    "_meta": { "title": "RandomNoise" }
+  },
+  "14": {
+    "inputs": {
+      "noise": ["13", 0],
+      "guider": ["10", 0],
+      "sampler": ["11", 0],
+      "sigmas": ["12", 0],
+      "latent_image": ["8", 2]
+    },
+    "class_type": "SamplerCustomAdvanced",
+    "_meta": { "title": "SamplerCustomAdvanced" }
+  },
+  "15": {
+    "inputs": {
+      "samples": ["14", 0],
+      "vae": ["3", 0]
+    },
+    "class_type": "VAEDecode",
+    "_meta": { "title": "VAE Decode" }
+  },
+  "16": {
+    "inputs": {
+      "frame_rate": 24,
+      "loop_count": 0,
+      "filename_prefix": "video/LTX23",
+      "format": "video/h264-mp4",
+      "pix_fmt": "yuv420p",
+      "crf": 19,
+      "save_metadata": true,
+      "trim_to_audio": false,
+      "pingpong": false,
+      "save_output": true,
+      "images": ["15", 0]
+    },
+    "class_type": "VHS_VideoCombine",
+    "_meta": { "title": "Save MP4" }
+  },
+  "20": {
+    "inputs": {
+      "clip_name1": "gemma-3-12b-it-qat-UD-Q4_K_XL.gguf",
+      "clip_name2": "text_encoders\\ltx-2.3-22b-dev_embeddings_connectors.safetensors",
+      "type": "ltxv"
+    },
+    "class_type": "DualCLIPLoaderGGUF",
+    "_meta": { "title": "DualCLIPLoader (GGUF)" }
+  }
+}

server/utils/comfy.ts

@@ -1,3 +1,5 @@
+import { LTX_NEGATIVE } from '~/utils/videoModels'
+
 let comfyHostOverride = ''
 
 export function setComfyHostOverride(url: string) {
@@ -188,7 +190,12 @@ export function extractPromptFromHistory(entry: unknown) {
     const text = String(node.inputs?.prompt || node.inputs?.text || '').trim()
     if (!text) continue
     if (node?.class_type === 'MiniMaxH3ReferenceToVideo') return text
-    if (node?.class_type === 'MiniMaxH3ImageToVideo') imageToVideo = text
+    if (node?.class_type === 'MiniMaxH3ImageToVideo' || node?.class_type === 'MiniMaxH3TextToVideo') imageToVideo = text
+    if (node?.class_type === 'CLIPTextEncode') {
+      const title = String((node as { _meta?: { title?: string } })._meta?.title || '')
+      if (/negative/i.test(title) || text === LTX_NEGATIVE) continue
+      if (!imageToVideo) imageToVideo = text
+    }
   }
   return imageToVideo
 }
@@ -200,14 +207,18 @@ export function extractClipMetaFromHistory(entry: unknown) {
   if (!graph || typeof graph !== 'object') return meta
   for (const node of Object.values(graph as Record<string, { class_type?: string; inputs?: Record<string, unknown> }>)) {
     const inputs = node?.inputs || {}
-    if (node?.class_type === 'MiniMaxH3ImageToVideo') {
+    if (node?.class_type === 'MiniMaxH3ImageToVideo' || node?.class_type === 'MiniMaxH3TextToVideo' || node?.class_type === 'EmptyLTXVLatentVideo') {
       meta.width = Number(inputs.width || meta.width)
       meta.height = Number(inputs.height || meta.height)
     }
-    if (node?.class_type === 'BasicScheduler' || node?.class_type === 'PrimitiveInt') {
+    if (node?.class_type === 'BasicScheduler' || node?.class_type === 'PrimitiveInt' || node?.class_type === 'LTXVScheduler') {
       const steps = Number(inputs.steps || 0)
       if (steps > 0) meta.steps = steps
     }
+    if (node?.class_type === 'ImageResize+' && typeof inputs.width === 'number' && typeof inputs.height === 'number') {
+      meta.width = Number(inputs.width)
+      meta.height = Number(inputs.height)
+    }
     if (node?.class_type === 'RandomNoise' || node?.class_type === 'SamplerMiniMax' || node?.class_type === 'KSampler') {
       const seed = Number(inputs.noise_seed ?? inputs.seed ?? 0)
       if (seed) meta.seed = seed
@@ -276,13 +287,16 @@ export function comfyFilenamePrefix() {
 }
 
 export function isOurComfyVideo(video: { filename: string; subfolder: string }) {
-  const prefix = comfyFilenamePrefix().replace(/\/$/, '')
-  const parts = prefix.split('/')
-  const namePrefix = parts[parts.length - 1]
-  const sub = parts.length > 1 ? parts.slice(0, -1).join('/') : 'video'
-  const nameOk = video.filename.startsWith(namePrefix)
-  const subOk = !video.subfolder || video.subfolder === sub
-  return nameOk && subOk
+  const prefixes = [comfyFilenamePrefix(), 'video/LTX23']
+  return prefixes.some((raw) => {
+    const prefix = raw.replace(/\/$/, '')
+    const parts = prefix.split('/')
+    const namePrefix = parts[parts.length - 1]
+    const sub = parts.length > 1 ? parts.slice(0, -1).join('/') : 'video'
+    const nameOk = video.filename.startsWith(namePrefix) || video.filename.startsWith('LTX23')
+    const subOk = !video.subfolder || video.subfolder === sub
+    return nameOk && subOk
+  })
 }
 
 export function purgeComfyEnabled() {

server/utils/jobs.ts

@@ -85,7 +85,7 @@ export interface Job {
     familyId?: string
     parentClipId?: string
     parentStillId?: string
-    workflow?: 'v1' | 'v2'
+    workflow?: import('~/utils/videoModels').VideoWorkflowId
     chainContinuing?: boolean
     passes?: { prompt: string }[]
     queueId?: string

server/utils/library.ts

@@ -38,7 +38,7 @@ export interface LibraryClip {
   familyId?: string
   parentClipId?: string
   chainIndex?: number
-  workflow?: 'v1' | 'v2'
+  workflow?: import('~/utils/videoModels').VideoWorkflowId
   stillId?: string
   sound?: boolean
   hideInput?: boolean
@@ -102,7 +102,7 @@ export interface RetryDraft {
   samplerName?: string
   scheduler?: string
   extensions?: QueuedExtension[]
-  workflow?: 'v1' | 'v2'
+  workflow?: import('~/utils/videoModels').VideoWorkflowId
 }
 
 interface Catalog {
@@ -969,7 +969,7 @@ export async function saveRetryDraft(params: {
   samplerName?: string
   scheduler?: string
   extensions?: QueuedExtension[]
-  workflow?: 'v1' | 'v2'
+  workflow?: import('~/utils/videoModels').VideoWorkflowId
 }) {
   return mutate(params.ownerKey, (catalog) => {
     const existing = params.id ? catalog.drafts.find(item => item.id === params.id) : undefined
@@ -1134,7 +1134,7 @@ export async function saveClip(params: {
   familyId?: string
   parentClipId?: string
   chainIndex?: number
-  workflow?: 'v1' | 'v2'
+  workflow?: import('~/utils/videoModels').VideoWorkflowId
   stillId?: string
   sound?: boolean
 }) {

server/utils/pending.ts

@@ -35,7 +35,7 @@ export interface PendingJob {
   referenceImageNames?: string[]
   referenceStillIds?: Array<string | null>
   useIdentityRefs?: boolean
-  workflow?: 'v1' | 'v2'
+  workflow?: import('~/utils/videoModels').VideoWorkflowId
   duration?: number
   cfg?: number
   fps?: number

server/utils/promptComfy.ts

@@ -1,5 +1,6 @@
 import { getSidecarImageHost } from '~/server/utils/imageComfy'
 import { buildVisionPromptWorkflow } from '~/server/utils/promptWorkflow'
+import { normalizeShotScript } from '~/utils/parseRecommend'
 
 export type PromptJobStatus = 'queued' | 'running' | 'complete' | 'error'
 
@@ -206,7 +207,10 @@ function draftUserPrompt(draft: string, mode: string) {
     : mode === 'shot-script'
       ? 'video-shot-script'
       : 'video'
-  return `Studio mode: ${studio}\n\nDraft Prompt:\n${draft}`
+  const format = studio === 'image-edit'
+    ? 'If this is a single edit, output only the refined prompt. If it needs phases, use ---pass 1--- blocks.'
+    : 'Output a MiniMax shot list. Each shot starts with "shot N" on its own line, then [SHOT CONFIGURATION], [SUBJECT DIRECTION & ACTION], and [AUDIO CUES], each on its own line. Blank line between shots. No markdown fences. No one-paragraph dump.'
+  return `Studio mode: ${studio}\n${format}\n\nDraft Prompt:\n${draft}`
 }
 
 export async function runPromptRecommend(job: PromptJob, params: {
@@ -227,7 +231,8 @@ export async function runPromptRecommend(job: PromptJob, params: {
   const queued = await queuePromptGraph(graph, job.id)
   job.message = 'Writing the recommended prompt…'
   const text = await waitForPromptText(queued.prompt_id)
-  job.text = text.trim()
+  const trimmed = text.trim()
+  job.text = params.mode === 'edit' ? trimmed : normalizeShotScript(trimmed)
   job.status = 'complete'
   job.message = 'Recommended prompt ready'
 }

server/utils/promptPresets.ts

@@ -9,6 +9,49 @@ export interface PromptPreset {
   builtin?: boolean
 }
 
+const SHOT_SCRIPT_FORMAT = `Output ONLY a MiniMax shot list the user can paste into Shot script mode.
+
+Each shot starts with "shot N" on its own line (shot 1, shot 2, shot 3…). Then these three headers, each on its own line:
+
+[SHOT CONFIGURATION]
+[SUBJECT DIRECTION & ACTION]
+[AUDIO CUES]
+
+Blank line between shots. No preamble, no markdown fences, no commentary.
+Do not glue shots into one paragraph. "shot 2" must be a line by itself, never "shot 2:" mid-sentence.
+
+One beat per shot. Later shots: same people, same clothes as now, plus the one new action.
+Do not repeat a long identity paragraph on every shot.
+Never censor raw or explicit terminology.
+
+Example:
+
+shot 1
+[SHOT CONFIGURATION]
+- Shot Type: Medium-wide, 50mm, eye-level
+- Camera Movement: Slow track from a slight side-front angle
+- Lighting: Bright daylight through the trees
+- Environment/Scenery: The forest from the still
+
+[SUBJECT DIRECTION & ACTION]
+The woman from the still walks the path, taking in the trees.
+
+[AUDIO CUES]
+Footsteps on dirt, birds, breeze.
+
+shot 2
+[SHOT CONFIGURATION]
+- Shot Type: Medium, 50mm, eye-level
+- Camera Movement: Hold, then a small pan to her hand
+- Lighting: Same daylight
+- Environment/Scenery: Same path, same trees
+
+[SUBJECT DIRECTION & ACTION]
+She stops and rests her hand on the bark.
+
+[AUDIO CUES]
+Soft fabric, breeze, birds.`
+
 const ENHANCE_PROMPT = `You are a prompt-engineering assistant for Flux.2 Klein image edit and MiniMax H3 image-to-video.
 
 Look at the attached still. The person (or subject) in that still is the identity lock. The user draft is the requested change.
@@ -16,37 +59,27 @@ Look at the attached still. The person (or subject) in that still is the identit
 Write production-ready prompt text the user can paste into the generator.
 
 Rules:
-- Preserve identity from the still: face, age, body type, hair, skin, and distinctive features. Use explicit keep-identity language such as "same person as the reference still", "do not change facial identity", "keep original face, hair, and body".
+- Preserve identity from the still: face, age, body type, hair, skin, and distinctive features.
 - Describe wardrobe, setting, pose, lighting, camera, and materials concretely.
 - Never censor raw or explicit terminology from the draft.
 - Do not add a preamble, title, markdown fences, or commentary.
 
-Output format — pick one:
+If Studio mode is image-edit:
 - Single still/edit: output ONLY the refined prompt.
 - Ambitious still edit that should be done in phases (for example wardrobe first, then location): output:
 ---pass 1---
 <prompt for pass 1>
 ---pass 2---
 <prompt for pass 2>
-- Motion, multi-beat, or video sequence: first-shot prompt, then later shots with a marker on its own line:
-shot 2
-<extension prompt>
-shot 3
-<extension prompt>`
+
+If Studio mode is video or video-shot-script:
+${SHOT_SCRIPT_FORMAT}`
 
 const SHOT_SCRIPT_PROMPT = `You are a MiniMax H3 image-to-video shot-script writer.
 
-Analyze the attached still for identity, wardrobe, lighting, and camera. Turn the user's draft into a shot script they can paste into Shot script mode.
+Analyze the attached still for wardrobe, lighting, and camera. Turn the user's draft into a shot script.
 
-Format:
-- Shot 1 is the text BEFORE any marker (camera, action, audio for the first clip).
-- Later shots start with a line that is only: shot 2
-- Then shot 3, shot 4, and so on, each on its own line.
-- Every shot must keep the same person as the still unless the draft explicitly changes wardrobe or identity.
-- Include an [AUDIO CUES] block on shots where sound would help.
-- Never censor raw or explicit terminology.
-
-Output ONLY the shot script. No preamble, no markdown fences.`
+${SHOT_SCRIPT_FORMAT}`
 
 export const BUILTIN_PRESETS: PromptPreset[] = [
   {
@@ -60,7 +93,7 @@ export const BUILTIN_PRESETS: PromptPreset[] = [
     id: 'shot-script',
     name: 'Shot script',
     builtin: true,
-    description: 'Turn a draft into MiniMax shot 2 / shot 3 extension copy.',
+    description: 'Turn a draft into a MiniMax shot list with configuration, action, and audio blocks.',
     systemPrompt: SHOT_SCRIPT_PROMPT
   }
 ]

server/utils/shotQueue.ts

@@ -38,7 +38,7 @@ export interface ShotQueue {
   fps?: number
   samplerName?: string
   scheduler?: string
-  workflow?: 'v1' | 'v2'
+  workflow?: import('~/utils/videoModels').VideoWorkflowId
   sound?: boolean
   useIdentityRefs?: boolean
   referenceStillIds?: Array<string | null>
@@ -153,7 +153,7 @@ export async function createShotQueue(params: {
   fps?: number
   samplerName?: string
   scheduler?: string
-  workflow?: 'v1' | 'v2'
+  workflow?: import('~/utils/videoModels').VideoWorkflowId
   sound?: boolean
   useIdentityRefs?: boolean
   referenceStillIds?: Array<string | null>

server/utils/videoChain.ts

@@ -5,6 +5,7 @@ import { pendingFromJob, remainingAfterCurrentShot, writePendingJob } from '~/se
 import { emitChainJob, waitForComfySocket, watchComfyJob } from '~/server/utils/watch'
 import { comfyFilenamePrefix, queuePrompt, uploadImage } from '~/server/utils/comfy'
 import { buildWorkflow } from '~/server/utils/workflow'
+import { isLtxWorkflow, isTextToVideo, parseVideoWorkflow, workflowForExtension, type VideoWorkflowId } from '~/utils/videoModels'
 import { clipVideoPath, clipTitle, deleteRetryDraft, extendTempDir, getClip, nextClipPartName, removeExtendTemp, stillPath } from '~/server/utils/library'
 import { extractLastFrame, probeHasAudio } from '~/server/utils/ffmpeg'
 import { ensureComfyReady } from '~/server/utils/comfyLifecycle'
@@ -23,7 +24,7 @@ export type ChainImage = { filename: string; data: Buffer; type?: string }
 
 type VideoChainParams = {
   prompt: string
-  image: ChainImage
+  image?: ChainImage | null
   width: number
   height: number
   steps: number
@@ -36,7 +37,7 @@ type VideoChainParams = {
   samplerName: string
   scheduler: string
   extensions: { prompt: string; duration: number }[]
-  workflow: 'v1' | 'v2'
+  workflow: VideoWorkflowId
   duration: number
   useIdentityRefs: boolean
   referenceImages: Array<ChainImage | null>
@@ -76,7 +77,8 @@ function paramsFromJob(job: Job): VideoChainParams {
     throw new Error('Cannot continue the shot chain: job metadata is missing')
   }
   const image = loadStillImage(library.ownerKey, library.stillId, library.stillFilename)
-  if (!image?.data.length) {
+  const workflow = parseVideoWorkflow(library.workflow)
+  if (!image?.data.length && !isTextToVideo(workflow)) {
     throw new Error('Cannot continue the shot chain: the start still is missing from the library')
   }
   const referenceImages: Array<ChainImage | null> = [null, null, null, null]
@@ -101,7 +103,7 @@ function paramsFromJob(job: Job): VideoChainParams {
     samplerName: library.samplerName || 'res_multistep',
     scheduler: library.scheduler || 'simple',
     extensions: library.extensions || [],
-    workflow: library.workflow || 'v1',
+    workflow,
     duration,
     useIdentityRefs: library.useIdentityRefs === true,
     referenceImages
@@ -121,14 +123,21 @@ export async function queueMiniMax(
   const done = watchComfyJob(job, { persist: params.persist })
   job.status = 'uploading'
   const chainIndex = job.library?.chainIndex || 0
-  const uploading = params.useIdentityRefs
-    ? (chainIndex > 0 ? 'Uploading identity stills for next shot...' : 'Uploading image to ComfyUI...')
-    : (chainIndex > 0 ? 'Uploading last frame to ComfyUI...' : 'Uploading image to ComfyUI...')
-  const queueing = (job.library?.chainIndex || 0) > 0
-    ? 'Queueing extension on MiniMax H3...'
-    : 'Queueing MiniMax H3 job...'
+  const graphId = chainIndex > 0 ? workflowForExtension(params.workflow) : params.workflow
+  const engineName = isLtxWorkflow(graphId) ? 'LTX-2.3' : 'MiniMax H3'
+  const hasImage = Boolean(params.image?.data?.length)
+  const uploading = !hasImage
+    ? `Queueing ${engineName} text-to-video…`
+    : params.useIdentityRefs
+      ? (chainIndex > 0 ? 'Uploading identity stills for next shot...' : 'Uploading image to ComfyUI...')
+      : (chainIndex > 0 ? 'Uploading last frame to ComfyUI...' : 'Uploading image to ComfyUI...')
+  const queueing = chainIndex > 0
+    ? `Queueing extension on ${engineName}...`
+    : `Queueing ${engineName} job...`
   emitChainJob(job, { type: 'status', message: uploading, progress: 4 })
-  const uploaded = await uploadImage(params.image, job.id)
+  const uploaded = hasImage && params.image
+    ? await uploadImage(params.image, job.id)
+    : { name: '', subfolder: '' }
   const referenceNames: string[] = ['', '', '', '']
   if (params.useIdentityRefs) {
     for (const [index, ref] of (params.referenceImages || []).entries()) {
@@ -161,9 +170,9 @@ export async function queueMiniMax(
     fps: params.fps,
     samplerName: params.samplerName,
     scheduler: params.scheduler,
-    filenamePrefix: comfyFilenamePrefix(),
-    sound: params.sound,
-    workflow: params.workflow,
+    filenamePrefix: isLtxWorkflow(graphId) ? 'video/LTX23' : comfyFilenamePrefix(),
+    sound: params.sound && !isLtxWorkflow(graphId),
+    workflow: graphId,
     duration: params.duration,
     useIdentityRefs: params.useIdentityRefs,
     referenceImageNames: params.useIdentityRefs ? referenceNames : []
@@ -317,7 +326,7 @@ export async function continueQueuedExtensions(
         samplerName: params.samplerName,
         scheduler: params.scheduler,
         persist,
-        workflow: params.workflow,
+        workflow: workflowForExtension(params.workflow),
         duration: ext.duration,
         useIdentityRefs: params.useIdentityRefs,
         referenceImages: params.useIdentityRefs ? params.referenceImages : []

server/utils/workflow.ts

@@ -1,9 +1,18 @@
 // Nitro bundles these JSON graphs into the production server output.
 import workflowV1 from '../assets/workflow_minimax_video.json'
 import workflowV2 from '../assets/workflow_minimax_video_v2.json'
+import workflowLtx from '../assets/workflow_ltx_video.json'
 import { buildIdentityPrompt } from '~/utils/identityPrompt'
+import {
+  isLtxWorkflow,
+  isTextToVideo,
+  LTX_NEGATIVE,
+  parseVideoWorkflow,
+  type VideoWorkflowId
+} from '~/utils/videoModels'
 
-export type VideoWorkflowId = 'v1' | 'v2'
+export type { VideoWorkflowId }
+export { parseVideoWorkflow }
 
 export interface GenerateParams {
   prompt: string
@@ -31,6 +40,7 @@ type WorkflowGraph = Record<string, WorkflowNode>
 
 const PROMPT_CLASSES = new Set([
   'MiniMaxH3ImageToVideo',
+  'MiniMaxH3TextToVideo',
   'MiniMaxH3ReferenceToVideo',
   'CLIPTextEncode',
   'CLIPTextEncodeQwen3VL',
@@ -58,16 +68,15 @@ function titleOf(node: WorkflowNode) {
 }
 
 function templateFor(id: VideoWorkflowId) {
+  if (isLtxWorkflow(id)) return workflowLtx as WorkflowGraph
   return (id === 'v2' ? workflowV2 : workflowV1) as WorkflowGraph
 }
 
-export function parseVideoWorkflow(raw: unknown): VideoWorkflowId {
-  return String(raw || '').trim() === 'v2' ? 'v2' : 'v1'
-}
-
 export function buildWorkflow(params: GenerateParams) {
-  const version = params.workflow === 'v2' ? 'v2' : 'v1'
-  const graph = structuredClone(templateFor(version))
+  const version = parseVideoWorkflow(params.workflow)
+  if (isLtxWorkflow(version)) return buildLtxWorkflow(params, version)
+
+  const graph = structuredClone(templateFor(version === 't2v' ? 'v1' : version === 'v2' ? 'v2' : 'v1'))
   const width = snap32(params.width)
   const height = snap32(params.height)
   const refs = Array.from({ length: 4 }, (_, index) => String(params.referenceImageNames?.[index] || '').trim())
@@ -155,7 +164,7 @@ export function buildWorkflow(params: GenerateParams) {
     }
   }
 
-  if (version === 'v1' && graph['128']?.class_type === 'ImageScaleToTotalPixels') {
+  if (version === 'v1' && !isTextToVideo(version) && graph['128']?.class_type === 'ImageScaleToTotalPixels') {
     graph['128'] = {
       class_type: 'ImageScale',
       inputs: {
@@ -169,7 +178,7 @@ export function buildWorkflow(params: GenerateParams) {
     }
   }
 
-  if (version === 'v1') {
+  if (version === 'v1' || version === 't2v') {
     graph['105:cfg'] = {
       class_type: 'FluxGuidance',
       inputs: {
@@ -191,10 +200,11 @@ export function buildWorkflow(params: GenerateParams) {
       graph[qualityId].inputs.value = params.turbo ? 20 : params.steps
     }
     if (turboId && graph[turboId]?.class_type === 'PrimitiveInt') {
-      graph[turboId].inputs.value = params.turbo ? params.steps : (params.workflow === 'v2' ? 6 : 8)
+      graph[turboId].inputs.value = params.turbo ? params.steps : (version === 'v2' ? 6 : 8)
     }
   }
 
+  if (version === 't2v') applyMinimaxTextToVideo(graph, params, width, height)
   if (version === 'v2') applyV2IdentityPath(graph, params, useRefs, refs, graphPrompt)
 
   if (params.sound === false) {
@@ -203,6 +213,7 @@ export function buildWorkflow(params: GenerateParams) {
     if (graph['105:91']?.inputs) delete graph['105:91'].inputs.audio
     delete graph['163']
     if (graph['172']?.inputs) delete graph['172'].inputs.audio
+    if (graph['105:104']?.inputs) delete graph['105:104'].inputs.audio_vae
   }
 
   return graph
@@ -210,10 +221,76 @@ export function buildWorkflow(params: GenerateParams) {
 
 function promptForNode(classType: string, useRefs: boolean, actionPrompt: string, graphPrompt: string) {
   if (classType === 'MiniMaxH3ImageToVideo') return useRefs ? '' : actionPrompt
+  if (classType === 'MiniMaxH3TextToVideo') return actionPrompt
   if (classType === 'MiniMaxH3ReferenceToVideo') return useRefs ? graphPrompt : actionPrompt
   return graphPrompt
 }
 
+function applyMinimaxTextToVideo(graph: WorkflowGraph, params: GenerateParams, width: number, height: number) {
+  const node = graph['105:104']
+  if (node) {
+    node.class_type = 'MiniMaxH3TextToVideo'
+    node._meta = { title: 'MiniMax H3 Text to Video' }
+    delete node.inputs.first_frame
+    node.inputs.prompt = params.prompt
+    node.inputs.width = width
+    node.inputs.height = height
+    if (params.sound !== false && graph['105:24']) node.inputs.audio_vae = ['105:24', 0]
+    else delete node.inputs.audio_vae
+  }
+  delete graph['114']
+  delete graph['119']
+  delete graph['120']
+  delete graph['127']
+  delete graph['128']
+}
+
+function buildLtxWorkflow(params: GenerateParams, version: VideoWorkflowId) {
+  const graph = structuredClone(workflowLtx as WorkflowGraph)
+  const width = snap32(params.width)
+  const height = snap32(params.height)
+  const textToVideo = isTextToVideo(version)
+  if (graph['4']?.inputs) graph['4'].inputs.image = params.imageName
+  if (graph['5']?.inputs) {
+    graph['5'].inputs.width = width
+    graph['5'].inputs.height = height
+  }
+  if (graph['6']?.inputs) graph['6'].inputs.text = params.prompt
+  if (graph['7']?.inputs) graph['7'].inputs.text = LTX_NEGATIVE
+  if (graph['8']?.inputs && typeof graph['8'].inputs.length === 'number') graph['8'].inputs.length = params.length
+  if (graph['9']?.inputs) graph['9'].inputs.frame_rate = params.fps
+  if (graph['10']?.inputs) graph['10'].inputs.cfg = params.cfg
+  if (graph['11']?.inputs) graph['11'].inputs.sampler_name = params.samplerName || 'euler'
+  if (graph['12']?.inputs) graph['12'].inputs.steps = params.steps
+  if (graph['13']?.inputs) graph['13'].inputs.noise_seed = params.seed
+  if (graph['16']?.inputs) {
+    graph['16'].inputs.frame_rate = params.fps
+    graph['16'].inputs.filename_prefix = params.filenamePrefix || 'video/LTX23'
+  }
+  if (graph['21']?.inputs) graph['21'].inputs.strength_model = 1
+  if (textToVideo) {
+    delete graph['4']
+    delete graph['5']
+    graph['8'] = {
+      class_type: 'EmptyLTXVLatentVideo',
+      inputs: {
+        width,
+        height,
+        length: params.length,
+        batch_size: 1
+      },
+      _meta: { title: 'Empty LTX latent' }
+    }
+    if (graph['9']?.inputs) {
+      graph['9'].inputs.positive = ['6', 0]
+      graph['9'].inputs.negative = ['7', 0]
+    }
+    if (graph['12']?.inputs) graph['12'].inputs.latent = ['8', 0]
+    if (graph['14']?.inputs) graph['14'].inputs.latent_image = ['8', 0]
+  }
+  return graph
+}
+
 function applyV2IdentityPath(
   graph: WorkflowGraph,
   params: GenerateParams,
@@ -304,9 +381,10 @@ function labelsFrom(template: WorkflowGraph) {
 export const NODE_LABELS: Record<string, string> = {
   ...Object.fromEntries(labelsFrom(workflowV1 as WorkflowGraph)),
   ...Object.fromEntries(labelsFrom(workflowV2 as WorkflowGraph)),
+  ...Object.fromEntries(labelsFrom(workflowLtx as WorkflowGraph)),
   ...LABEL_OVERRIDES
 }
 
 export function isEncodingNode(node: string) {
-  return node === '92' || node === '105:91' || node === '172' || /encoding|saving mp4|create video/i.test(NODE_LABELS[node] || '')
+  return node === '92' || node === '16' || node === '105:91' || node === '172' || /encoding|saving mp4|create video|save mp4/i.test(NODE_LABELS[node] || '')
 }