Add image-to-text Describe on the bench via exclusive llama.cpp caption jobs.

2026-09-30 22:50:19 -05:00
parent 9481791ac5
commit 15ef891b26
15 changed files with 937 additions and 35 deletions
@@ -234,6 +234,18 @@
      >
       <button type="button" @click="applyUseImage()">Use image</button>
       <button type="button" @click="applyUsePrompt()">Use prompt</button>
+       <button
+        v-if="selected.kind==='still'"
+        type="button"
+        :disabled="submitting"
+        @click="describeSelected"
+       >Describe</button>
+       <label v-if="selected.kind==='still'" class="inline-flex items-center gap-1 text-[var(--type-sm)] text-[var(--muted)]">
+        <span class="sr-only">Caption style</span>
+        <select v-model="captionStyle" class="field-control !w-auto !py-1 !text-[var(--type-sm)]" aria-label="Caption style">
+         <option v-for="style in captionStyles" :key="style" :value="style">{{ style }}</option>
+        </select>
+       </label>
       <div class="stage-split">
        <button
         type="button"
@@ -382,6 +394,11 @@
       />
       <span v-if="card.elapsed != null" class="job-elapsed">{{ card.elapsed }}</span>
      </div>
+      <p v-if="card.resultText" class="mt-2 whitespace-pre-wrap text-[var(--type-sm)]">{{ card.resultText }}</p>
+      <div v-if="card.resultText" class="mt-2 flex flex-wrap gap-2">
+       <button type="button" class="text-[var(--type-sm)] text-[var(--amber)]" @click.stop="copyCaption(card.resultText)">Copy</button>
+       <button type="button" class="text-[var(--type-sm)] text-[var(--amber)]" @click.stop="useCaptionAsPrompt(card.resultText)">Use as prompt</button>
+      </div>
     </div>
     <p v-if="!queueCards.length" class="px-1 text-[var(--type-sm)] text-[var(--muted)]">No active jobs</p>
    </div>
@@ -460,6 +477,7 @@ import {aspectOptions,resolveSize,bindClip,restoreAll,useAsSource,hydrateStudio2
 import {folderAssets,newestOutput,assetLabel,reconcileSelection,previewHidden} from '~/shared/studio2/viewer.mjs'
 import {queueJobs,jobStageLabel,jobProgressView,elapsedSeconds} from '~/shared/studio2/queue-view.mjs'
 import {gpuStatus} from '~/shared/studio2/status.mjs'
+import {CAPTION_STYLES} from '~/shared/caption.mjs'
 import {applyUploadedStill} from '~/shared/studio2/upload.mjs'
 import {defaultNegatives,parseBatch} from '~/shared/studio2/styles.mjs'
 import {randomSeed} from '~/shared/studio2/seed.mjs'
@@ -614,6 +632,9 @@ const promptArea=ref<HTMLTextAreaElement|null>(null)
 const promptCaret=ref({start:0,end:0})
 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 liveCaption=ref<Record<string,any>>({})
 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))
@@ -725,7 +746,10 @@ const benchJobs=computed(()=>queueJobs(folderJobs.value))
 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 captionJobs=computed(()=>studioQueue.value
+ .filter((j:any)=>j.kind==='caption' && j.folderId===form.folderId)
+ .map((j:any)=>({...j,live:j.liveJobId?liveCaption.value[j.liveJobId]:null})))
+const gpu=computed(()=>gpuStatus(jobs.value,healthState.value,upscales.value,musicJobs.value,captionJobs.value))
 const gpuChipLabel=computed(()=>{
 if(!healthState.value) return 'GPU ?'
 const raw=String(healthState.value?.comfy?.gpu || healthState.value?.gpuName || healthState.value?.comfy?.device || '').trim()
@@ -792,6 +816,22 @@ const queueCards=computed(()=>{
   cancelable:!['complete','done','error','cancelled'].includes(job.status)
  })
 }
+ for(const job of captionJobs.value){
+  const live=job.live
+  const text=live?.resultText || job.resultText || ''
+  const state=live?.status || job.status
+  cards.push({
+   id:job.id,
+   job,
+   code:'#'+String(job.id).replace(/\D/g,'').slice(-3).padStart(3,'0'),
+   title:job.name || `Describe · ${job.captionStyle || job.payload?.captionStyle || 'caption'}`,
+   meta:`Caption · ${job.captionStyle || job.payload?.captionStyle || 'descriptive'}`,
+   state,
+   statusLabel:live?.message || state,
+   resultText:text,
+   cancelable:!['complete','done','error','cancelled'].includes(job.status) && !['complete','error','cancelled'].includes(String(state))
+  })
+ }
 for(const job of upscales.value.filter((j:any)=>j.folderId===form.folderId && ['queued','running'].includes(j.status))){
  cards.push({
   id:job.id,
@@ -807,7 +847,7 @@ const queueCards=computed(()=>{
 return cards
 })
 const queueCount=computed(()=>queueCards.value.length)
-const canClearFinished=computed(()=>folderJobs.value.some(j=>['complete','failed','cancelled'].includes(j.state)) || musicJobs.value.some(j=>['complete','done','error','cancelled'].includes(j.status)))
+const canClearFinished=computed(()=>folderJobs.value.some(j=>['complete','failed','cancelled'].includes(j.state)) || musicJobs.value.some(j=>['complete','done','error','cancelled'].includes(j.status)) || captionJobs.value.some(j=>['complete','done','error','cancelled'].includes(j.status) || ['complete','error','cancelled'].includes(j.live?.status)))
 const transportLabel=computed(()=>{
 if(selected.value?.kind==='clip') return `${formatTime(videoTime.value)} / ${formatTime(videoDuration.value)}`
 if(selected.value?.kind==='track') return 'Audio'
@@ -1000,6 +1040,23 @@ function applyUseImage(item?:any){
 uploadMessage.value=`Added ${assetLabel(asset)} to Photos.`
 saveBenchDraft(true)
 }
+async function copyCaption(text:string){
+ try{await navigator.clipboard.writeText(text)}catch{error.value='Could not copy caption.'}
+}
+function useCaptionAsPrompt(text:string){
+ const current=String(form.promptSections.action||'').trim()
+ form.promptSections.action=current?`${current}\n\n${text}`:text
+}
+async function describeSelected(){
+ const id=selected.value?.kind==='still'?selected.value.id:''
+ if(!id){error.value='Select a still to describe.';return}
+ submitting.value=true;error.value=''
+ try{
+  await $fetch('/api/generate/caption',{method:'POST',body:{folderId:form.folderId,stillId:id,sourcePath:id,captionStyle:captionStyle.value}})
+  await refreshJobs()
+ }catch(e:any){error.value=e.data?.statusMessage||e.message}
+ finally{submitting.value=false}
+}
 function applyUsePrompt(item?:any){
 const asset=resolveActionAsset(item)
 if(!asset){
@@ -1369,6 +1426,14 @@ async function refreshJobs(){
 const nextLive:Record<string,any>={}
 for(const [id,snap] of liveEntries){if(snap)nextLive[id]=snap}
 liveMusic.value=nextLive
+ const captionRows=studioQueue.value.filter((j:any)=>j.kind==='caption' && j.liveJobId)
+ const captionEntries=await Promise.all(captionRows.map(async(j:any)=>{
+  try{return [j.liveJobId,await $fetch<any>(`/api/generate/${j.liveJobId}`)] as const}
+  catch{return [j.liveJobId,null] as const}
+ }))
+ const nextCaption:Record<string,any>={}
+ for(const [id,snap] of captionEntries){if(snap)nextCaption[id]=snap}
+ liveCaption.value=nextCaption
 ingestNewOutputs(nextJobs)
 }
 /** Merge freshly completed outputs into the local catalog — never refetch the whole library. */
@@ -1559,7 +1624,14 @@ async function queue(){
  logTiming('done')
 }
 }
-async function cancel(job:any){await $fetch('/api/studio-2/cancel',{method:'POST',body:{id:job.id}});await refreshJobs()}
+async function cancel(job:any){
+ if(job?.kind==='music' || job?.kind==='caption'){
+  await $fetch(`/api/studio-queue/${job.id}`,{method:'DELETE'})
+ }else{
+  await $fetch('/api/studio-2/cancel',{method:'POST',body:{id:job.id}})
+ }
+ await refreshJobs()
+}

 async function loadLibraryOnce(){
 // Current folder page only — never the yearly Inbox JSON on /
@@ -0,0 +1,299 @@
+import { spawn } from 'node:child_process'
+import { createServer } from 'node:net'
+import { existsSync, mkdirSync, readFileSync, writeFileSync, renameSync, appendFileSync, readdirSync, copyFileSync, createWriteStream } from 'node:fs'
+import { join, resolve, extname } from 'node:path'
+import { pipeline } from 'node:stream/promises'
+import { Transform } from 'node:stream'
+import { captionStylePrompt, CAPTION_STYLES } from '../shared/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'))
+}
+
+/** Prefer winget ggml.llamacpp; never download. Process exit after one caption = keep_alive 0. */
+function resolveLlamaServer(explicit) {
+  if (explicit) return resolve(explicit)
+  if (process.env.CAPTION_LLAMA_SERVER) return resolve(process.env.CAPTION_LLAMA_SERVER)
+  const local = process.env.LOCALAPPDATA || ''
+  const candidates = [
+    join(local, 'Microsoft', 'WinGet', 'Packages', 'ggml.llamacpp_Microsoft.Winget.Source_8wekyb3d8bbwe', 'llama-server.exe'),
+    join(local, 'llama.cpp', 'llama-server.exe'),
+    join('C:\\', 'llama.cpp', 'llama-server.exe')
+  ]
+  for (const path of candidates) {
+    if (existsSync(path)) return path
+  }
+  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:${port}/health`, { signal: AbortSignal.timeout(2000) })
+      if (response.ok) return
+    } catch { /* booting */ }
+    await new Promise(r => setTimeout(r, 500))
+  }
+  throw new Error('llama-server did not become ready in time.')
+}
+
+function mimeFor(path) {
+  const ext = extname(path).toLowerCase()
+  if (ext === '.jpg' || ext === '.jpeg') return 'image/jpeg'
+  if (ext === '.webp') return 'image/webp'
+  if (ext === '.gif') return 'image/gif'
+  return 'image/png'
+}
+
+function killTree(child) {
+  return new Promise(resolveKill => {
+    if (!child?.pid) return resolveKill()
+    let finished = false
+    const done = () => { if (finished) return; finished = true; resolveKill() }
+    child.once('close', done)
+    try { if (typeof child.kill === 'function') child.kill() } catch { /* ignore */ }
+    if (process.platform === 'win32' && child.pid > 0) {
+      try {
+        const killer = spawn('taskkill', ['/PID', String(child.pid), '/T', '/F'], { windowsHide: true, stdio: 'ignore' })
+        killer.once('close', done)
+        killer.once('error', done)
+      } catch { done() }
+    }
+    setTimeout(done, 3000)
+  })
+}
+
+export function validateCaptionHostRequest(body) {
+  if (!/^[a-zA-Z0-9-]{12,80}$/.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 unloads VRAM (keep_alive 0). */
+export function createCaptionHost({
+  prepare,
+  leaseValid,
+  spawnProcess = spawn,
+  modelsDir,
+  llamaServer,
+  dataDir,
+  fetchImpl = fetch,
+  serverWaitMs = 120_000
+} = {}) {
+  const models = resolve(modelsDir || defaultModelsDir())
+  const executable = resolveLlamaServer(llamaServer)
+  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 */ }
+    }
+  }
+  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 }
+    const path = join(dir(id), 'status.json')
+    return existsSync(path) ? JSON.parse(readFileSync(path, 'utf8')) : null
+  }
+  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')
+      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 missing 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 = existsSync(dir(request.id)) && 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()
+        // One caption per process. killTree after response unloads VRAM (keep_alive 0).
+        const child = spawnProcess(executable, [
+          '-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'
+        ], { 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, serverWaitMs, 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 response = await fetchImpl(`http://127.0.0.1:${port}/v1/chat/completions`, {
+            method: 'POST',
+            headers: { 'Content-Type': 'application/json' },
+            signal: run.abort.signal.aborted ? run.abort.signal : AbortSignal.timeout(Number(process.env.CAPTION_TIMEOUT_MS || 180_000)),
+            body: JSON.stringify({
+              temperature: 0.2,
+              max_tokens: 512,
+              messages: [{
+                role: 'user',
+                content: [
+                  { type: 'text', text: captionStylePrompt(request.style) },
+                  { type: 'image_url', image_url: { url: dataUrl } }
+                ]
+              }]
+            })
+          })
+          if (run.abort.signal.aborted) throw new Error('Cancelled')
+          if (!response.ok) {
+            const detail = await response.text().catch(() => '')
+            throw new Error(`llama-server caption failed (${response.status}): ${detail.slice(0, 400)}`)
+          }
+          const payload = await response.json()
+          const text = String(payload?.choices?.[0]?.message?.content || '').trim()
+          if (!text) throw new Error('Caption model returned empty text.')
+          job.text = text
+          job.status = run.cancelled ? 'cancelled' : 'complete'
+          job.message = job.status === 'complete' ? 'Caption ready' : 'Cancelled'
+          job.progress = 100
+          job.checkedAt = Date.now()
+          persist(job)
+          return { ...job }
+        } finally {
+          clearInterval(watchdog)
+          await killTree(child)
+          holdUntil = Date.now() + 2000
+          if (active === run) active = null
+        }
+      } catch (error) {
+        job.status = run.cancelled ? 'cancelled' : 'error'
+        job.error = error.message || String(error)
+        job.message = job.error
+        job.checkedAt = Date.now()
+        persist(job)
+        if (run.child) await killTree(run.child)
+        if (active === run) active = null
+        holdUntil = Date.now() + 2000
+        throw error
+      }
+    },
+    async cancel(id) {
+      if (active?.job.id !== id) return read(id)
+      const run = active
+      run.cancelled = true
+      run.job.status = 'cancelling'
+      run.job.message = 'Cancelling'
+      persist(run.job)
+      run.abort.abort()
+      if (run.child) await killTree(run.child)
+      return read(id)
+    },
+    async captionOnce(body, lease) {
+      const id = body.id || `caption-${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 }
@@ -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) {
@@ -830,29 +831,65 @@ const gpuReservation = createGpuReservation({ idle: async () => {
  return !(await processUp()) && !(await pythonMainUp().catch(() => true))
 } })

-const yue2 = createYue2Host({
-  leaseValid: lease => gpuReservation.isOwner(lease),
-  prepare: async () => {
+async function prepareExclusiveGpu(label) {
  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)) {
    await stopComfyProcesses()
    markAsleep()
  }
  if (await pythonMainUp()) throw new Error('Comfy has not stopped; retry after the GPU is free.')
-  }
-})
+}

 const upscale = createUpscaleHost({ leaseValid: token => gpuReservation.isOwner(token) })

+const yue2 = createYue2Host({
+  leaseValid: lease => gpuReservation.isOwner(lease),
+  prepare: async () => 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)?$/)
+    if (req.method === 'GET' && url.pathname === '/caption/status') {
+      return json(res, 200, { configured: caption.configured(), busy: caption.busy(), backend: 'llama.cpp', modelsDir: caption.modelsDir() })
+    }
+    if (req.method === 'POST' && url.pathname === '/caption') {
+      try {
+        const body = await readJson(req)
+        return json(res, 200, await caption.captionOnce(body, String(req.headers['x-aigen-gpu-lease'] || '')))
+      } catch (error) {
+        return json(res, error.statusCode || 400, { error: error.message || 'Caption failed' })
+      }
+    }
+    if (req.method === 'POST' && url.pathname === '/caption/jobs') {
+      return json(res, 200, await caption.start(await readJson(req), String(req.headers['x-aigen-gpu-lease'] || '')))
+    }
+    if (match && req.method === 'PUT' && match[2] === '/input') {
+      await caption.upload(match[1], req)
+      return json(res, 200, { uploaded: true })
+    }
+    if (match && req.method === 'POST' && match[2] === '/cancel') return json(res, 200, await caption.cancel(match[1]))
+    if (match && req.method === 'GET' && !match[2]) {
+      const job = caption.read(match[1])
+      return json(res, job ? 200 : 404, job || { error: 'Caption job not found' })
+    }
+    return json(res, 404, { error: 'Unknown caption endpoint' })
+  }
  if (url.pathname.startsWith('/upscale/')) {
    const match = url.pathname.match(/^\/upscale\/jobs\/([a-zA-Z0-9-]{12,80})(\/input|\/video|\/cancel)?$/)
    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 +946,14 @@ 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 +1055,7 @@ 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.' })
  }
 })
@@ -0,0 +1,54 @@
+# Caption (Qwen2.5-VL NSFW Caption V4 GGUF) — verify host files only.
+# Models are already on the 5080 host. Do not download. Do not search Hugging Face.
+
+$ErrorActionPreference = 'Stop'
+
+function Resolve-ModelsRoot {
+  if ($env:COMFY_MODELS_ROOT) { return $env:COMFY_MODELS_ROOT }
+  $shared = Join-Path $env:LOCALAPPDATA 'Comfy-Desktop\ComfyUI-Shared\models'
+  if (Test-Path $shared) { return $shared }
+  throw 'Set COMFY_MODELS_ROOT or install Comfy Desktop Shared models.'
+}
+
+$root = Join-Path (Resolve-ModelsRoot) 'caption\qwen25vl-7b-nsfw-v4'
+if (-not (Test-Path $root)) {
+  throw "Caption models folder missing: $root"
+}
+
+$files = @(
+  'Qwen2.5-VL-7B-NSFW-Caption-V4.Q5_K_M.gguf',
+  'Qwen2.5-VL-7B-NSFW-Caption-V4.mmproj-f16.gguf'
+)
+
+foreach ($name in $files) {
+  $path = Join-Path $root $name
+  if (-not (Test-Path $path)) {
+    throw "Missing $name under $root. Do not download; use the files already on this host."
+  }
+  Write-Host "OK $name"
+}
+
+$llama = $null
+if ($env:CAPTION_LLAMA_SERVER -and (Test-Path $env:CAPTION_LLAMA_SERVER)) {
+  $llama = Get-Item $env:CAPTION_LLAMA_SERVER
+} else {
+  $winget = Join-Path $env:LOCALAPPDATA 'Microsoft\WinGet\Packages\ggml.llamacpp_Microsoft.Winget.Source_8wekyb3d8bbwe\llama-server.exe'
+  if (Test-Path $winget) { $llama = Get-Item $winget }
+  else { $llama = Get-Command llama-server -ErrorAction SilentlyContinue }
+}
+
+if (-not $llama) {
+  throw 'llama-server not found. Install with: winget install ggml.llamacpp'
+}
+$llamaPath = if ($llama.PSObject.Properties['FullName'] -and $llama.FullName) { $llama.FullName } elseif ($llama.PSObject.Properties['Source'] -and $llama.Source) { $llama.Source } else { [string]$llama }
+Write-Host "llama-server: $llamaPath"
+
+$ready = @{
+  model = 'Qwen2.5-VL-7B-NSFW-Caption-V4.Q5_K_M.gguf'
+  mmproj = 'Qwen2.5-VL-7B-NSFW-Caption-V4.mmproj-f16.gguf'
+  modelsDir = $root
+  llamaServer = $llamaPath
+  readyAt = (Get-Date).ToUniversalTime().ToString('o')
+} | ConvertTo-Json
+Set-Content -Path (Join-Path $root 'aigen-ready.json') -Value $ready -Encoding utf8
+Write-Host "Caption ready: $root"
@@ -18,6 +18,12 @@ Get-NetTCPConnection -LocalPort 8199 -State Listen -ErrorAction SilentlyContinue

 Start-Sleep -Milliseconds 400

+# Caption VLM — point at existing host GGUFs (no download).
+$captionModels = Join-Path $env:LOCALAPPDATA 'Comfy-Desktop\ComfyUI-Shared\models\caption\qwen25vl-7b-nsfw-v4'
+$llamaServer = Join-Path $env:LOCALAPPDATA 'Microsoft\WinGet\Packages\ggml.llamacpp_Microsoft.Winget.Source_8wekyb3d8bbwe\llama-server.exe'
+if (Test-Path $captionModels) { $env:CAPTION_MODELS = $captionModels }
+if (Test-Path $llamaServer) { $env:CAPTION_LLAMA_SERVER = $llamaServer }
+
 if ($Wait) {
  Set-Location $root
  & $node $agent
@@ -31,4 +37,5 @@ $psi.WorkingDirectory = $root
 $psi.UseShellExecute = $false
 $psi.CreateNoWindow = $true
 $psi.WindowStyle = [System.Diagnostics.ProcessWindowStyle]::Hidden
+# CAPTION_* already set on this process; child inherits when UseShellExecute is false
 [void][System.Diagnostics.Process]::Start($psi)
@@ -0,0 +1,50 @@
+import { validateCaptionBody } from '~/shared/caption.mjs'
+import { addStudioJob, kickStudioQueue, type StudioJobPayload } from '~/server/utils/studioQueue'
+import { getStill } from '~/server/utils/library'
+
+export default defineEventHandler(async (event) => {
+  const { owner } = assertLibraryOwner(event)
+  const raw = await readBody(event)
+  let body: ReturnType<typeof validateCaptionBody>
+  try {
+    body = validateCaptionBody(raw)
+  } catch (e: any) {
+    throw createError({ statusCode: e.statusCode || 400, statusMessage: e.message })
+  }
+  assertFolderAccess(event, body.folderId)
+  const still = getStill(owner, body.stillId)
+  assertFolderAccess(event, still.folderId)
+
+  const row = await addStudioJob({
+    ownerKey: owner,
+    kind: 'caption',
+    familyId: crypto.randomUUID(),
+    payload: {
+      prompt: `Describe · ${body.captionStyle}`,
+      name: `Describe · ${body.captionStyle}`,
+      folderId: body.folderId,
+      stillId: still.id,
+      captionStyle: body.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
+  })
+  kickStudioQueue()
+  return { id: row.id, liveJobId: row.liveJobId, captionStyle: body.captionStyle, stillId: still.id }
+})
@@ -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,
@@ -10,13 +10,14 @@ export default defineEventHandler((event) => {
      || job.status === 'running'
      || job.status === 'held'
      || (job.status === 'error' && Date.now() - job.updatedAt < 60 * 60 * 1000)
+      || (job.kind === 'caption' && job.status === 'complete' && Date.now() - job.updatedAt < 60 * 60 * 1000)
    ))
  const claimed = new Set(active.map(job => job.shotQueueId).filter(Boolean) as string[])
  const queues = listShotQueues(owner)
  const byId = new Map(queues.map(queue => [queue.id, queue]))
  const jobs = active.map((job) => {
    const shots = job.shotQueueId ? byId.get(job.shotQueueId) || null : null
-    const kind = job.kind === 'edit' ? 'edit' : job.kind === 'music' ? 'music' : 'video'
+    const kind = job.kind === 'edit' ? 'edit' : job.kind === 'music' ? 'music' : job.kind === 'caption' ? 'caption' : 'video'
    const shotSummary = shots ? summarizeQueue(shots) : null
    const plannedShots = shotSummary
      ? undefined
@@ -0,0 +1,3 @@
+import { resumeCaptionJobs } from '~/server/utils/caption'
+
+export default defineNitroPlugin(() => { resumeCaptionJobs() })
@@ -0,0 +1,232 @@
+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/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) {
+  return Boolean(jobId) && 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 } : {}),
+    // Caption loads a 7B VLM; allow the host POST to finish or fall through to watch
+    signal: AbortSignal.timeout(method === 'PUT' ? 120_000 : method === 'POST' ? 300_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 */ }
+}
+
+export async function cancelCaptionJob(job: Job) {
+  await request(`jobs/${job.id}/cancel`, 'POST', {})
+  job.status = 'cancelled'
+  emitJob(job, { type: 'error', error: 'Cancelled', message: 'Cancelled' })
+}
+
+async function settle(job: Job) {
+  const { onLiveVideoSettled } = await import('./studioQueue')
+  await onLiveVideoSettled(job)
+  clearPending(job.id)
+}
+
+async function watch(job: Job, record: {
+  id: string
+  owner: string
+  stillId: string
+  folderId: string
+  captionStyle: string
+  liveId: string
+}) {
+  let failures = 0
+  while (job.status !== 'cancelled') {
+    try {
+      const state = await (await request(`jobs/${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 })
+        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.resultText = text
+        writeSidecar(record.owner, record.stillId, text)
+        persist({ ...record, resultText: text, status: 'complete' })
+        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
+  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 · ${params.captionStyle}`,
+    prompt: params.captionStyle,
+    aspect: 'image',
+    width: 0,
+    height: 0,
+    steps: 1,
+    turbo: true,
+    seed: 0,
+    stillId: params.stillId,
+    engine: 'caption'
+  }
+  const record = {
+    id: job.id,
+    liveId: job.id,
+    owner: params.ownerKey,
+    stillId: params.stillId,
+    folderId: params.folderId,
+    captionStyle: params.captionStyle,
+    createdAt: Date.now(),
+    status: 'running'
+  }
+  persist(record)
+  void (async () => {
+    try {
+      if (job.status === 'cancelled') { await settle(job); return }
+      await request(`jobs/${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 on 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 })
+        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'
+      if (record.resultText) job.resultText = record.resultText
+      job.message = 'Reconnecting to caption host'
+      void watch(job, record)
+    } catch { /* keep bad records */ }
+  }
+}
@@ -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
@@ -222,6 +224,7 @@ export function jobSnapshot(job: Job) {
    clipId: job.clipId,
    stillId: job.stillId,
    trackId: job.trackId,
+    resultText: job.resultText,
    hideThumbnail: job.hideThumbnail,
    error: job.error,
    folderLocked: job.library?.folderLocked,
@@ -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
@@ -42,6 +42,7 @@ export interface StudioJobPayload {
  useIdentityRefs: boolean
  stillId?: string
  stillFilename?: string
+  captionStyle?: string
  hideThumbnail: boolean
  hideInput?: boolean
  folderLocked?: boolean
@@ -105,6 +106,7 @@ export interface StudioJob {
  resumeAutoRun?: boolean
  lastError?: string
  waitReason?: string
+  resultText?: string
 }

 type StudioQueueStore = {
@@ -226,6 +228,7 @@ 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'
 }

@@ -245,6 +248,8 @@ export function summarizeStudioJob(job: StudioJob) {
    folderId: job.payload.folderId,
    musicEngine: job.payload.musicEngine,
    stillId: job.payload.stillId,
+    captionStyle: job.payload.captionStyle,
+    resultText: job.resultText,
    workflow: job.payload.workflow,
    imagePipeline: job.payload.imagePipeline || 'v1',
    duration: job.payload.duration,
@@ -290,7 +295,7 @@ 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')
@@ -310,12 +315,12 @@ 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')
@@ -355,7 +360,7 @@ 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
@@ -375,7 +380,7 @@ 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') {
@@ -469,8 +474,11 @@ export async function addStudioJob(params: {
  kind?: StudioJobKind
 }) {
  const now = Date.now()
-  const kind = params.kind === 'edit' ? 'edit' : params.kind === 'music' ? 'music' : 'video'
-  const shotCount = kind === 'music'
+  const kind = params.kind === 'edit' ? 'edit'
+    : params.kind === 'music' ? 'music'
+      : params.kind === 'caption' ? 'caption'
+        : 'video'
+  const shotCount = kind === 'music' || kind === 'caption'
    ? 1
    : kind === 'edit'
      ? 1 + (params.payload.passes?.length || 0)
@@ -606,6 +614,10 @@ 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) {
@@ -678,6 +690,11 @@ 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) {
@@ -889,6 +906,10 @@ 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 => (
@@ -1397,6 +1418,32 @@ 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,
+      name: item.payload.name
+    })
+    await markStudioLive(item.ownerKey, item.id, live.id)
+  } 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 {
@@ -1439,6 +1486,10 @@ 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')
@@ -1703,7 +1754,7 @@ 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) => {
@@ -1777,6 +1828,7 @@ export async function onLiveVideoSettled(job: Job) {
      : failed
        ? 'error'
        : 'complete'
+    if (job.caption && job.resultText) row.resultText = job.resultText
    clearStudioRowSlot(row, status, failed ? job.error : undefined)
    syncPausedFlag(store)
  })
@@ -0,0 +1,30 @@
+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 validateCaptionBody(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.sourcePath || '')
+  if (!stillId) fail('Choose a still to describe.')
+  return { folderId, stillId, captionStyle: style }
+}
@@ -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','running','held'].includes(j.status) || ['queued','running','uploading'].includes(j.live?.status))
+ if(caption)return {label:caption.status==='waiting'?'Queued':'Describing',tone:'busy',detail:caption.payload?.captionStyle || caption.live?.message || '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','enhancing','submitting','rendering','saving'].includes(j.state))
 if(job)return {label:{waking:'Waking up',enhancing:'Enhancing prompt…',submitting:'Submitting',rendering:'Generating…',saving:'Saving'}[job.state],tone:'busy',detail:job.request?.promptSections?.action || 'Generation in progress'}
@@ -0,0 +1,52 @@
+import assert from 'node:assert/strict'
+import { createCaptionHost, validateCaptionHostRequest } from '../scripts/caption-host.mjs'
+import { CAPTION_STYLES, captionStylePrompt, DELTA_LOCKED_LOOK, validateCaptionBody } from '../shared/caption.mjs'
+import { mkdtempSync, writeFileSync, mkdirSync, existsSync } from 'node:fs'
+import { join } from 'node:path'
+import { tmpdir } from 'node:os'
+import { EventEmitter } from 'node:events'
+
+assert.deepEqual(CAPTION_STYLES, ['descriptive', 'klein_prompt', 'delta', 'tags'])
+assert.match(captionStylePrompt('delta'), new RegExp(DELTA_LOCKED_LOOK.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')))
+assert.equal(validateCaptionBody({ folderId: 'f1', stillId: 's1', captionStyle: 'tags' }).captionStyle, 'tags')
+assert.throws(() => validateCaptionBody({ folderId: 'f1' }), /still/)
+assert.throws(() => validateCaptionHostRequest({ id: 'short', style: 'descriptive' }), /Invalid/)
+
+const root = mkdtempSync(join(tmpdir(), 'caption-test-'))
+const models = join(root, 'models')
+mkdirSync(models)
+writeFileSync(join(models, 'Qwen2.5-VL-7B-NSFW-Caption-V4.Q5_K_M.gguf'), 'x')
+writeFileSync(join(models, 'Qwen2.5-VL-7B-NSFW-Caption-V4.mmproj-f16.gguf'), 'x')
+writeFileSync(join(root, 'llama-server.exe'), 'x')
+
+let prepared = 0
+const host = createCaptionHost({
+  modelsDir: models,
+  llamaServer: join(root, 'llama-server.exe'),
+  dataDir: join(root, 'jobs'),
+  leaseValid: () => true,
+  prepare: async () => { prepared += 1 },
+  serverWaitMs: 800,
+  spawnProcess: () => {
+    const child = new EventEmitter()
+    child.pid = 4242
+    child.stdout = new EventEmitter()
+    child.stderr = new EventEmitter()
+    child.kill = () => child.emit('close', 0)
+    setTimeout(() => child.emit('close', 0), 50)
+    return child
+  },
+  fetchImpl: async () => ({ ok: false, status: 503, text: async () => 'down' })
+})
+
+assert.equal(host.configured(), true)
+assert.equal(host.busy(), false)
+
+const id = 'caption-test-abcdefghijkl'
+mkdirSync(join(root, 'jobs', id), { recursive: true })
+writeFileSync(join(root, 'jobs', id, 'input.png'), Buffer.alloc(64, 1))
+
+await assert.rejects(() => host.start({ id, style: 'descriptive' }, 'lease'), /llama-server|Caption/)
+assert.equal(prepared, 1)
+assert.ok(existsSync(join(root, 'jobs', id, 'status.json')))
+console.log('caption tests ok')