dsh-cosplay-studio 2.0.4 → 2.1.0

package/lib/index.js

@@ -8,7 +8,7 @@
 //   4. post-processing (de-oil / JPEG cap / 2×2 preview / delivery copy);
 //   5. the JSON API the browser half talks to, plus project media serving.
 
-import { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs'
+import { existsSync, mkdirSync, readFileSync, readdirSync, writeFileSync } from 'node:fs'
 import { join, relative, resolve, sep } from 'node:path'
 import {
   API_PREFIX, MEDIA_PREFIX, PROJECTS_DIR, dshHome, logLine, patchConfig, readConfig, setupLog, clearLog,
@@ -18,13 +18,15 @@ import {
   modelStatus, resolveComfyLayout, selfTest, startComfy, stopComfy, waitForComfy,
 } from './comfy.js'
 import {
-  buildCard, defaultDetails, llmConfigured, researchCharacter, ruleScript, shotPrompt, writeScript,
+  SCRIPT_SYSTEM, buildCard, chatComplete, defaultDetails, llmConfigured, normalizeShotItems, researchCharacter,
+  ruleScript, scriptUserText, shotPrompt, writeScript,
 } from './card.js'
 import {
-  cancelQueue, createProject, enqueueBase, enqueueFinalize, enqueueReferenceBase, enqueueShots,
+  cancelQueue, createProject, enqueueBase, enqueueEdit, enqueueFinalize, enqueueShots,
   listProjects, projectDirs, queueSnapshot, readProject, saveProject, updateProject,
 } from './project.js'
 import { recommend } from './hot.js'
+import { createRandomProject, llmText, parseJson, researchState, startResearch } from './research.js'
 
 const inject = ['webServer']
 
@@ -44,11 +46,17 @@ const API = {
   update: `${API_PREFIX}/project/update`,
   reference: `${API_PREFIX}/project/reference`,
   base: `${API_PREFIX}/project/base`,
+  baseEdit: `${API_PREFIX}/project/base/edit`,
+  shotEdit: `${API_PREFIX}/project/shot/edit`,
+  restoreShot: `${API_PREFIX}/project/shot/restore`,
   script: `${API_PREFIX}/project/script`,
+  scriptAdd: `${API_PREFIX}/project/script/add`,
   generate: `${API_PREFIX}/project/generate`,
   finalize: `${API_PREFIX}/project/finalize`,
   queue: `${API_PREFIX}/queue`,
   openFolder: `${API_PREFIX}/project/open-folder`,
+  research: `${API_PREFIX}/project/research`,
+  random: `${API_PREFIX}/project/random`,
 }
 
 const TUTORIAL = [
@@ -280,7 +288,7 @@ function referenceSave(id, body) {
 function applyShots(id, shots) {
   const project = readProject(id)
   if (project === null) throw new Error('项目不存在')
-  const normalized = (Array.isArray(shots) ? shots : []).slice(0, 6).map((shot, index) => ({
+  const normalized = (Array.isArray(shots) ? shots : []).slice(0, 10).map((shot, index) => ({
     id: index + 1,
     type: String(shot.type ?? 'A'),
     title: String(shot.title ?? `镜头 ${index + 1}`),
@@ -295,6 +303,81 @@ function applyShots(id, shots) {
   return updateProject(id, { shots: normalized })
 }
 
+function mediaTypeOfName(name) {
+  const lower = String(name).toLowerCase()
+  if (lower.endsWith('.png')) return 'image/png'
+  if (lower.endsWith('.webp')) return 'image/webp'
+  return 'image/jpeg'
+}
+
+/** Script model path: configured chat endpoint first, then the DSH host model,
+ *  then the built-in templates. Keeps existing shots for smart additions. */
+async function generateShotList(ctx, project, count, existing = []) {
+  const config = readConfig()
+  const user = scriptUserText(project.details, count, existing)
+  const startIndex = existing.length + 1
+  if (llmConfigured(config)) {
+    try {
+      const text = await chatComplete(config, SCRIPT_SYSTEM, user)
+      const parsed = parseJson(text)
+      const list = Array.isArray(parsed) ? parsed : parsed?.shots
+      return { shots: normalizeShotItems(list, startIndex), note: '镜头脚本由脚本模型生成' }
+    } catch {
+      // fall through to the host model
+    }
+  }
+  try {
+    let lastError = null
+    for (const maxTokens of [9000, 12000]) {
+      try {
+        const text = await llmText(ctx, `${SCRIPT_SYSTEM}\n\n${user}\n\n只输出 JSON 数组，不要输出任何思考过程或解释。`, { maxTokens })
+        const parsed = parseJson(text)
+        const list = Array.isArray(parsed) ? parsed : parsed?.shots
+        return { shots: normalizeShotItems(list, startIndex), note: '镜头脚本由 DSH 模型生成' }
+      } catch (error) {
+        lastError = error
+      }
+    }
+    throw lastError ?? new Error('脚本模型没有返回内容')
+  } catch (error) {
+    return { shots: ruleScript(project.details, Math.max(1, Math.min(6, count))), note: `内置脚本模板（${error?.message ?? error}）` }
+  }
+}
+
+/** Smart edit: compare the picture with the reference / approved base, let the
+ *  vision model name the flaws and produce one English img2img instruction. */
+async function smartEditInstruction(ctx, project, target, file, note = '') {
+  const dirs = projectDirs(project.id)
+  const images = []
+  const add = (folder, name) => {
+    if (name === '') return
+    const path = join(folder, name)
+    if (!existsSync(path) || images.some((image) => image.name === name)) return
+    images.push({ data: readFileSync(path), mediaType: mediaTypeOfName(name), name })
+  }
+  add(target === 'base' ? dirs.base : dirs.raw, file)
+  add(dirs.refs, String(project.referenceFile ?? ''))
+  add(dirs.base, String(project.approvedBase ?? ''))
+  const ask = note === ''
+    ? '没有额外要求，请自行找出最需要修的瑕疵（手、脚、五官、服装边缘、穿模）'
+    : `用户的修改要求：${note}`
+  const prompt = `你在为一套真人 cosplay 照片做精修指导，只输出 JSON。\n\n角色设定（英文）：\n${String(project.card ?? '').slice(0, 1800)}\n\n${ask}\n\n请比对图片：第一张是待修改的图片${images.length > 1 ? '，后面的图片是该角色的参考图 / 已确认基准图' : ''}。找出需要修的 1-4 处具体问题，并给出一句可直接用于 img2img 重绘的英文编辑指令：只描述要改的局部，明确保持同一人物、同一姿势、同一机位、同一服装与背景不变；优先修手、脚、五官和衣物边缘这类细节。\n\n只输出：{"issues":["中文问题1","中文问题2"],"fix":"one English editing instruction"}`
+  let lastError = null
+  for (const maxTokens of [4000, 6000]) {
+    try {
+      const text = await llmText(ctx, `${prompt}\n\n不要输出任何思考过程。`, { images, maxTokens })
+      const parsed = parseJson(text)
+      const issues = Array.isArray(parsed?.issues) ? parsed.issues.map((value) => String(value)).filter(Boolean).slice(0, 4) : []
+      const fix = String(parsed?.fix ?? '').trim()
+      if (fix === '') throw new Error('智能分析没有给出修改指令')
+      return { issues, fix }
+    } catch (error) {
+      lastError = error
+    }
+  }
+  throw lastError ?? new Error('智能分析失败')
+}
+
 /* ----------------------------------------------------------------- apply */
 
 export function apply(ctx) {
@@ -342,6 +425,9 @@ export function apply(ctx) {
       updatedAt: project.updatedAt,
       finalCount: (project.finals ?? []).length,
       baseCount: (project.baseImages ?? []).length,
+      research: project.research
+        ? { status: project.research.status, stage: project.research.stage, stageLabel: project.research.stageLabel }
+        : null,
     }))
     sendJson(response, 200, {
       ok: true,
@@ -499,7 +585,10 @@ export function apply(ctx) {
       sendJson(response, 404, { ok: false, error: '项目不存在' })
       return
     }
-    sendJson(response, 200, { ok: true, project })
+    // A host restart leaves research jobs marked running forever; normalise so
+    // the panel shows "重新研究" instead of a spinner that never ends.
+    const research = researchState(project)
+    sendJson(response, 200, { ok: true, project: research === null ? project : { ...project, research } })
   })
 
   route(API.recommend, async (request, response) => {
@@ -535,6 +624,7 @@ export function apply(ctx) {
       return
     }
     const series = String(body?.series ?? '').trim()
+    const source = String(body?.source ?? 'name')
     const config = readConfig()
     let details = body?.details !== null && typeof body?.details === 'object' ? { ...body.details } : null
     let note = ''
@@ -544,17 +634,30 @@ export function apply(ctx) {
           details = await researchCharacter(config, { character, series, referencePath: body?.referenceFile ?? '' })
           note = '角色卡由脚本模型生成'
         } catch (error) {
-          details = defaultDetails(character, series)
+          details = defaultDetails(character, series, { reference: source === 'image' })
           note = `脚本模型不可用（${error?.message ?? error}），已用内置模板`
         }
       } else {
-        details = defaultDetails(character, series)
-        note = llmConfigured(config) ? '按内置模板生成角色卡' : '未配置脚本模型，按内置模板生成角色卡'
+        details = defaultDetails(character, series, { reference: source === 'image' })
+        note = source === 'image'
+          ? '参考图项目：人物外观以参考图为准，角色卡里的发型/服装/道具只是占位描述，可留空'
+          : llmConfigured(config) ? '按内置模板生成角色卡' : '未配置脚本模型，角色卡用通用模板；想做得像请上传参考图'
       }
     }
     const project = createProject({ character, series, details, source: body?.source ?? 'name' })
     updateProject(project.id, { notes: [note] })
-    sendJson(response, 200, { ok: true, project: readProject(project.id), note })
+    // Name projects start the slow research right away (web search → pictures →
+    // vision analysis → draft). Image projects start after the upload arrives.
+    let researching = false
+    if (body?.research !== false && source !== 'image') {
+      try {
+        startResearch(ctx, config, project.id)
+        researching = true
+      } catch (error) {
+        updateProject(project.id, { notes: [note, `角色研究未能启动：${error?.message ?? error}`] })
+      }
+    }
+    sendJson(response, 200, { ok: true, project: readProject(project.id), note, researching })
   })
 
   route(API.update, async (request, response) => {
@@ -579,7 +682,14 @@ export function apply(ctx) {
     if (Array.isArray(body?.shots)) {
       applyShots(id, body.shots)
     }
-    if (typeof body?.status === 'string') patch.status = body.status
+    if (typeof body?.status === 'string') {
+      patch.status = body.status
+      // Remember which base the user approved: later shots are img2img from it.
+      if (body.status === 'base-approved') {
+        const bases = Array.isArray(project.baseImages) ? project.baseImages : []
+        if (bases.length > 0) patch.approvedBase = bases[bases.length - 1]
+      }
+    }
     if (Object.keys(patch).length > 0) updateProject(id, patch)
     sendJson(response, 200, { ok: true, project: readProject(id) })
   })
@@ -628,19 +738,22 @@ export function apply(ctx) {
       sendJson(response, 409, { ok: false, error: '正在生成其它任务，请稍候' })
       return
     }
-    const useReference = body?.useReference === true && String(project.referenceFile ?? '') !== ''
-    if (useReference) enqueueReferenceBase(config, layout, id)
-    else enqueueBase(config, layout, id, body?.count === 2 ? 2 : 1)
+    // Reference projects default to img2img from the uploaded picture; the
+    // text-to-image path stays available with an explicit useReference:false.
+    const useReference = String(project.referenceFile ?? '') !== '' && body?.useReference !== false
+    enqueueBase(config, layout, id, body?.count === 2 ? 2 : 1, { useReference })
     sendJson(response, 200, { ok: true, queue: queueSnapshot() })
   })
 
-  route(API.script, async (request, response) => {
+  // Manual / smart retouch: 'base' appends a new base version, 'shot' replaces
+  // one raw final. Smart mode asks the vision model what to fix first.
+  const editRoute = (target) => async (request, response) => {
     if (request.method !== 'POST') {
       response.writeHead(405, { allow: 'POST' })
       response.end()
       return
     }
-    const body = await readBody(request, 1024 * 1024)
+    const body = await readBody(request, 64 * 1024)
     const id = String(body?.id ?? '')
     const project = readProject(id)
     if (project === null) {
@@ -648,26 +761,194 @@ export function apply(ctx) {
       return
     }
     const config = readConfig()
-    const count = Math.max(4, Math.min(6, Number(body?.count ?? config.shotCount ?? 6)))
-    let shots = null
-    let note = ''
-    if (body?.useModel !== false && llmConfigured(config)) {
+    const { layout } = detectComfyDir(config)
+    if (layout === null || (await comfyStats(config.comfyPort)) === null) {
+      sendJson(response, 409, { ok: false, error: 'ComfyUI 没有运行，请先在「安装」页启动' })
+      return
+    }
+    if (queueSnapshot().running) {
+      sendJson(response, 409, { ok: false, error: '正在生成其它任务，请稍候' })
+      return
+    }
+    const file = String(body?.file ?? '').trim()
+    if (target === 'shot' && file === '') {
+      sendJson(response, 400, { ok: false, error: '请选择要修改的成片' })
+      return
+    }
+    let instruction = String(body?.instruction ?? '').trim()
+    const notes = []
+    if (body?.smart === true) {
       try {
-        shots = await writeScript(config, project.details, count)
-        note = '镜头脚本由脚本模型生成'
+        const analysis = await smartEditInstruction(ctx, project, target, file, instruction)
+        instruction = analysis.fix
+        if (analysis.issues.length > 0) notes.push(`智能微调分析：${analysis.issues.join('；')}`)
       } catch (error) {
-        shots = ruleScript(project.details, count)
-        note = `脚本模型不可用（${error?.message ?? error}），已用内置脚本`
+        sendJson(response, 500, { ok: false, error: `智能分析失败：${error?.message ?? error}` })
+        return
       }
-    } else {
-      shots = ruleScript(project.details, count)
-      note = '内置脚本模板'
     }
+    if (instruction === '') {
+      sendJson(response, 400, { ok: false, error: '请输入修改要求，或改用智能修改' })
+      return
+    }
+    try {
+      enqueueEdit(config, layout, id, { target, file, instruction, denoise: Number(body?.denoise), notes })
+      sendJson(response, 200, { ok: true, project: readProject(id), issues: notes })
+    } catch (error) {
+      sendJson(response, 500, { ok: false, error: error?.message ?? String(error) })
+    }
+  }
+  route(API.baseEdit, editRoute('base'))
+  route(API.shotEdit, editRoute('shot'))
+
+  // Undo the most recent retouch of one raw final (backups live in raw/.history).
+  route(API.restoreShot, async (request, response) => {
+    if (request.method !== 'POST') {
+      response.writeHead(405, { allow: 'POST' })
+      response.end()
+      return
+    }
+    const body = await readBody(request, 16 * 1024)
+    const id = String(body?.id ?? '')
+    const project = readProject(id)
+    if (project === null) {
+      sendJson(response, 404, { ok: false, error: '项目不存在' })
+      return
+    }
+    const file = String(body?.file ?? '').trim()
+    if (file === '' || file.includes('/') || file.includes('\\') || file.includes('..')) {
+      sendJson(response, 400, { ok: false, error: '文件名不合法' })
+      return
+    }
+    const dirs = projectDirs(id)
+    const history = join(dirs.raw, '.history')
+    const current = join(dirs.raw, file)
+    if (!existsSync(history) || !existsSync(current)) {
+      sendJson(response, 404, { ok: false, error: '没有可恢复的历史版本' })
+      return
+    }
+    const versions = readdirSync(history)
+      .filter((name) => name.startsWith(`${file}.`) && name.endsWith('.png'))
+      .sort()
+    if (versions.length === 0) {
+      sendJson(response, 404, { ok: false, error: '没有可恢复的历史版本' })
+      return
+    }
+    const latest = versions[versions.length - 1]
+    writeFileSync(join(history, `${file}.${Date.now()}.png`), readFileSync(current))
+    writeFileSync(current, readFileSync(join(history, latest)))
+    const updated = updateProject(id, {
+      pendingFinals: [...new Set([...(project.pendingFinals ?? []), file])],
+      notes: [...(project.notes ?? []), `已恢复上一版成片：${file}`],
+    })
+    sendJson(response, 200, { ok: true, project: updated })
+  })
+
+  route(API.research, async (request, response) => {
+    if (request.method !== 'POST') {
+      response.writeHead(405, { allow: 'POST' })
+      response.end()
+      return
+    }
+    const body = await readBody(request, 64 * 1024)
+    const id = String(body?.id ?? '')
+    if (readProject(id) === null) {
+      sendJson(response, 404, { ok: false, error: '项目不存在' })
+      return
+    }
+    const config = readConfig()
+    try {
+      const project = startResearch(ctx, config, id, { supplement: body?.supplement === true })
+      sendJson(response, 200, { ok: true, project })
+    } catch (error) {
+      sendJson(response, 500, { ok: false, error: error?.message ?? String(error) })
+    }
+  })
+
+  route(API.random, async (request, response) => {
+    if (request.method !== 'POST') {
+      response.writeHead(405, { allow: 'POST' })
+      response.end()
+      return
+    }
+    const config = readConfig()
+    try {
+      const pick = await createRandomProject(ctx, config)
+      const project = createProject({
+        character: pick.character,
+        series: pick.series,
+        details: defaultDetails(pick.character, pick.series),
+        source: 'random',
+      })
+      updateProject(project.id, { notes: [pick.reason === '' ? '随机推荐角色' : `随机推荐：${pick.reason}`] })
+      startResearch(ctx, config, project.id)
+      sendJson(response, 200, { ok: true, project: readProject(project.id), pick })
+    } catch (error) {
+      sendJson(response, 500, { ok: false, error: error?.message ?? String(error) })
+    }
+  })
+
+  route(API.script, async (request, response) => {
+    if (request.method !== 'POST') {
+      response.writeHead(405, { allow: 'POST' })
+      response.end()
+      return
+    }
+    const body = await readBody(request, 1024 * 1024)
+    const id = String(body?.id ?? '')
+    const project = readProject(id)
+    if (project === null) {
+      sendJson(response, 404, { ok: false, error: '项目不存在' })
+      return
+    }
+    const config = readConfig()
+    const count = Math.max(4, Math.min(6, Number(body?.count ?? config.shotCount ?? 6)))
+    const { shots, note } = await generateShotList(ctx, project, count)
     const updated = applyShots(id, shots)
     updateProject(id, { status: 'awaiting-script-approval', notes: [...(updated.notes ?? []), note] })
     sendJson(response, 200, { ok: true, project: readProject(id), note })
   })
 
+  // Smart add: the script model designs extra shots that must not repeat the
+  // ones already in the list.
+  route(API.scriptAdd, async (request, response) => {
+    if (request.method !== 'POST') {
+      response.writeHead(405, { allow: 'POST' })
+      response.end()
+      return
+    }
+    const body = await readBody(request, 64 * 1024)
+    const id = String(body?.id ?? '')
+    const project = readProject(id)
+    if (project === null) {
+      sendJson(response, 404, { ok: false, error: '项目不存在' })
+      return
+    }
+    const existing = Array.isArray(project.shots) ? project.shots : []
+    if (existing.length === 0) {
+      sendJson(response, 409, { ok: false, error: '还没有镜头脚本，请先生成' })
+      return
+    }
+    const add = Math.max(1, Math.min(3, Number(body?.count ?? 2)))
+    const limit = Math.max(0, 10 - existing.length)
+    if (limit === 0) {
+      sendJson(response, 409, { ok: false, error: '已经有 10 个分镜，先删除一些再补' })
+      return
+    }
+    const { shots: fresh, note } = await generateShotList(ctx, project, Math.min(add, limit), existing)
+    const merged = [...existing, ...fresh].slice(0, 10).map((shot, index) => ({
+      ...shot,
+      id: index + 1,
+      prompt: `${shot.type ?? 'A'}｜${shot.title ?? `镜头 ${index + 1}`}`,
+    }))
+    const updated = updateProject(id, {
+      shots: merged,
+      status: 'awaiting-script-approval',
+      notes: [...(project.notes ?? []), note, `智能补充分镜 ${fresh.length} 个`],
+    })
+    sendJson(response, 200, { ok: true, project: updated, note })
+  })
+
   route(API.generate, async (request, response) => {
     if (request.method !== 'POST') {
       response.writeHead(405, { allow: 'POST' })
@@ -718,7 +999,13 @@ export function apply(ctx) {
       sendJson(response, 409, { ok: false, error: '已有任务在运行' })
       return
     }
-    enqueueFinalize(config, layout, id)
+    enqueueFinalize(config, layout, id, {
+      deoil: body?.deoil === true,
+      makeup: body?.makeup === true,
+      flatten: body?.flatten !== false,
+      preview: body?.preview !== false,
+      copy: body?.copy !== false,
+    })
     sendJson(response, 200, { ok: true })
   })