wind-comic

Wind Comic _^v13.17

One sentence in. A finished short-form drama out — script, cast, storyboards, voiceover, timeline, mp4.
Multi-agent AI studio · reusable characters · novel→season splitting · director's control room · real-time collab · bring-your-own LLM.

一句话进,整片短剧出 —— 剧本 · 角色 · 分镜 · 配音 · 时间线 · mp4 一条龙。
多 Agent AI 创作工作室 · 可复用角色 · 长篇小说→自动分集 · 导演级控片台 · 实时协作 · 自带 LLM。

English · 简体中文 · 🔥 Pitch · 🔌 BYO LLM

_{Mirror for mainland China · 国内 AtomGit 托管(自动同步自 GitHub):atomgit.com/ChrisChen667788/wind-comic}

▶ Watch the full 39-second promo — with voiceover & sound
_{Real cinematic footage woven with motion-graphics · 8 distinct art styles · English narration · scored by the platform's own MiniMax music engine.}

💎 Sponsors · 赞助商

MiniMax H3 视频生成 API｜秘塔科技 秘塔科技提供高性价比的 MiniMax H3 视频生成服务：768P 仅 0.09 元/秒，2K 仅 0.15 元/秒。支持原生 2K、音画同步，API 兼容 OpenAI 协议，同时支持 ComfyUI，无需自行部署 GPU。 🎁 通过 wind-comic专属链接注册，即可领取赠送额度及专属优惠。 MiniMax H3 video-generation API — MetaSota (秘塔科技). ¥0.09/s at 768P, ¥0.15/s at 2K · native 2K with synced audio · OpenAI-compatible API · ComfyUI supported · no GPU to deploy. 🎁 Sign up through the wind-comic link for bonus credits and a partner discount.

✨ Why Wind Comic?

Most “AI video” tools give you a 5-second clip from a one-line prompt. Wind Comic gives you a finished short-form drama — script, character bible, multi-shot storyboards, voice-acted lines, BGM, mouth-synced characters (indicative 2D lip bar out of the box; photorealistic face-level lip-sync via BYO wav2lip/SadTalker/MuseTalk or Kling), and a final mp4 — from the same single line.

It works because it doesn’t try to be one giant model. It’s an honest multi-agent pipeline where each role (Writer, Director, Producer, Character Designer, Storyboard Artist, Cameo Locker, Lipsync, Editor) is a specialist that hands off with strict consistency contracts. Plus a real-time multiplayer timeline to edit with your team like Figma for film.

   "A reborn CEO confronts his cheating ex-fiancée at her wedding."
                                       │
                                       ▼
   Writer ▶ Director ▶ Style Bible ▶ Char Designer ▶ Scene Designer ▶
   ▶ Storyboard (vision-audited) ▶ Video (multi-engine race) ▶
   ▶ TTS (per-character voice) ▶ Lipsync (Kling/Sync.so/Hailuo) ▶
   ▶ Editor (j-cut/l-cut + BGM per act + CJK subtitles) ▶ final.mp4

   + Real-time collab timeline (Yjs CRDT)
   + Bring-your-own LLM (3 env vars, 0 code change)
   + Plug-in image/video providers (12+ supported)
   + Lip-sync pipeline: viseme track → align score → drift auto-correct → engine render → back into timeline   (v10)
   + Template market: save a hit project → rate / favorite / one-click remix (with its voices)               (v10)
   + Per-project cost attribution + budget guard, four-dimension publish gate                                 (v10)

🏗 Architecture

Three views of the same engine. Open on GitHub to watch them animate — flowing dashes trace live data & control paths, traveling dots are data packets moving through the pipeline.

_{System architecture — five layers, top to bottom. The Director threads control across the seven pipeline agents (eight roles including the Director); the LLM gateway falls back DeepSeek → MiniMax with zero code change; 12+ media engines plug in behind one router; everything lands on a dual-driver (SQLite ⇄ PostgreSQL) platform.}

_{Sequence — the lifecycle of one idea → film request, time flowing down. Two signature beats: the Vision-Audit retry loop (auto-regenerate any shot scoring < 70) and the multi-engine race (Seedance / Kling / Veo / Vidu — first good clip wins).}

_{Data flow — the artifact refinery. One line of text is refined stage-by-stage (TEXT → JSON → PNG → IMG → MP4); every artifact is persisted to the dual-driver DB + asset store and is independently reusable, so any stage can be re-run in isolation.}

🎞️ Diagrams are animated SVG, authored as code in assets/diagrams/ — crisp at any zoom, versioned with the source. (Animation renders on GitHub; the ModelScope mirror shows static PNG.)

🆕 New in v13.0 — the frame you set is the frame you get

v13.0 consolidates v12.461 – v12.473 (13 releases, 2026-09-30 → 10-09). One pattern ran through most of them: settings that looked like they worked but had no code reading them, and engine requests that looked like they succeeded but sent the wrong thing.

Aspect ratio, end to end (v12.464, 467, 468, 470 – 473) — until v12.467 only full-film generation read the project’s aspect ratio. Regenerating a single shot, self-healing retakes, segment retake, 4K re-render, recomposing after a chat edit, scene-image regeneration and the ad workshop each fell back to their own 16:9 or 9:16, so a 9:16 project could get a landscape shot back. They now all read projects.aspect through one shared parser. Prompt templates no longer hard-code --ar 16:9 (a vertical project used to send Midjourney both --ar 16:9 and --ar 9:16); Midjourney parameters are moved to the end and appear once, and non-Midjourney image engines and the video engines receive plain text. The create page no longer offers 2.35:1, which no video engine we ship can output; the ad workshop’s aspect is now selectable and defaults to the project’s.

Settings that finally have a reader (v12.464, 466, 469, 473) — the format bar’s color space now goes into storyboard prompts, its safe-area toggle drives the 9:16 overlay, and its aspect is a read-only display of the project’s real aspect instead of a second, dead copy. The short-video planner dropped the resolution / frame-rate / upscaling / interpolation / motion-strength controls that nothing read (the defaults claimed 8K, 4× upscale and interpolation), and its aspect choice now carries over to the create page.

Director’s console, finished (v12.462, 463, 465) — add, remove and rename actors; the camera can tilt, and “aim at subjects” points a raised camera at the actors; the layout sketch re-renders when the blocking changes, and shots generated from the old blocking are marked stale. The sketch goes only to engines that compose from it (fal Kontext, Seedream; Midjourney as an opt-in) — never to MiniMax, whose reference input is portrait-only. The full walkthrough — real WebGL (software-rendered in CI) and real MediaPipe pose inference — now runs in CI.

Camera raised to 2.6 m, then aim at subjects: it tilts 19° down and the console now calls it a high-angle shot. The 3D view, the flat preview, the prompt line and the layout sketch all use this one camera.	Same camera, flat preview: the horizon sits high in the frame because the lens points down. Before v12.465 a raised camera still looked level, while the prompt already said "high angle".

Engine requests that were silently wrong (v12.461, 465, 471, 472) — MiniMax’s subject_reference[].image_file had been sent as an array since v2.20 while the official schema is a string, so every MiniMax generation with reference images was rejected (2013) and quietly fell back to another engine; a live call found it. The Vidu host default existed twice, one copy wrong (that copy had no reader, so no request ever used it — removed before anyone wired it up); Midjourney prompts carried duplicated and mid-prompt parameters; video engines were handed Midjourney syntax as prompt text.

Not verified yet: sketch delivery to Seedream / kontext / Midjourney follows their documented request formats but could not be confirmed with a live generation (gateway quota exhausted / no Midjourney channel / upstream saturated on 10-01, 10-06 and 10-08); npx tsx scripts/probe-sketch-engines.ts --yes <engine> re-runs the check, one paid image per engine. fal is unverified (no key configured).

🆕 New in v12.315 – v12.330 — direct the shot, don’t re-roll the dice

Every tool can generate a clip. These three let you direct one — and they compose into a single loop: inspect → locate → retake only what’s broken.

Director’s console (v12.316–318) — place the actors and the camera on a top-down stage; the exact blocking becomes two things a model can actually use: a precise staging directive appended to the prompt ("Lin Wan at frame left in full shot; Lu Chen right of center in wide shot"), and a layout sketch fed through the existing [STORYBOARD LOCK] channel (layout only — style still comes from the prompt; the sketch goes only to engines that actually compose from it — fal Kontext and Seedream, with Midjourney image prompts as an opt-in; MiniMax’s reference input is portrait-only, so it never gets the sketch. When no such engine is available the lock is skipped, and if generation falls back to an engine that didn’t receive the sketch the lock line is stripped — never claimed). Shot size and camera angle are derived from the geometry, not typed in by hand, so lens/framing and the written spec can’t disagree; the camera can tilt up and down (v12.465), and “aim at subjects” points a raised camera at the actors in one click — the prompt, the 3D preview and the layout sketch all describe the same camera. A deterministic composition audit reports who is out of frame, who is occluding whom, whether the camera clips an actor — before you spend a cent generating.

Segment retake (v12.315) — hate two seconds of an eight-second shot? Retake those two. The other six aren’t regenerated: they’re re-encoded once, in the same pass that splices in the new two seconds, at x264 crf 17 (the visually-lossless range). It’s not a byte copy — -c copy can only cut on keyframes, and here the cut has to land on the exact frame you picked. The whole audio track keeps the source’s sample rate and channel layout, so the spliced seconds don’t shift pitch. Shot duration is unchanged by construction, which means the compressed timeline, voice-over delays, subtitle starts and EDL record-ins need no recomputation. Takes are versioned like voice retakes — adopt or roll back. End to end since v12.459: box the bad frames in the frame inspector → a free dry-run shows what will be generated → confirm, and the server generates the patch starting from the source frame at the cut-in point, splices it, checks the video frame count before recording the take, and lists it for preview. The first adoption keeps the original as a take of its own, so rolling back is one click. If every video engine fails it refuses instead of splicing in a still-image placeholder, and the request is synchronous (engines take ~30–120 s).

Frame-by-frame inspection (v12.328–330) — step through a finished shot frame by frame, box the broken stretch, and hand that exact range to segment retake. Frames are extracted with accurate seek (-ss after -i, not the fast keyframe-only seek), and the timestamp under each frame uses the same frame-snapping as the retake planner — so the frame you picked is the frame it cuts at. When the strip is thinned to stay responsive, it says so; frames that fail to decode are reported, not silently skipped.

Also in this range: fonts are self-hosted so builds no longer depend on a network fetch; the director paths finally receive the target language (non-Chinese projects used to get Chinese scripts and rely on a costly after-the-fact re-translation); the in-process rate limiter bounds and evicts its bucket table (the key contained an attacker-supplied email — an unbounded memory vector), and eviction never releases an active block; collaborative comments arriving over Yjs are validated and field-whitelisted, so a peer can no longer overwrite another author’s name or text.

🆕 New in v12 — the quality loop era

Sketch-Lock storyboarding — every shot first renders a B/W composition sketch, then the final frame is generated locked to that composition (toggle at creation time, or per-shot in the Shot Workshop). Camera language finally survives the diffusion lottery.

Emotion-driven camera — shots missing camera movement get one derived from story beats (hook→push-in, chase→handheld, reveal→zoom-in, farewell→pull-out, emotion temperature → push/orbit/crane). Even the Ken Burns degradation follows the intended move.

Engine matrix, honestly — Veo · MiniMax (Hailuo) · Kling (official Beijing API) · HappyHorse 1.1 (Alibaba, v12.272 — joint video+audio in one pass) with a one-line env priority (VIDEO_ENGINE_ORDER=kling,minimax,veo,happyhorse), base64 first-frame channel for local assets, and an honest fallback chain: a failed shot becomes a labeled animatic, never a still image masquerading as video.

The quality loop — ffprobe-powered film health report (aspect/duration/fps/bitrate/audio/degraded shots) per project and per series; batch re-render of degraded shots (2-way concurrent); a season-export health gate so no broken episode sneaks into the compilation; an engine weather bar so you know before you hit generate.

Round-trip storyboards — export the pull sheet as CSV / Markdown / print-ready PDF (with character sheet + health appendix), edit in Excel, import it back — changed visual fields offer one-click re-render of exactly the affected shots.

Writer resilience — truncated or quote-mangled LLM scripts are salvaged (bracket-stack completion + content-quote escaping) instead of being silently replaced by a template. The real 23KB incident dump lives in the test fixtures.

🆕 Earlier: v6 → v10 — from demo to production platform

v3 shipped the pipeline. v6 turned it into a production studio; v7–v9 hardened it into a platform; v10 closed the lip-sync, template-market, and cost loops. Reusable characters, a prompt IDE, novel→season auto-splitting with real voiceover, a 60-style gallery, a director’s control room, team credit budgets, an industry-grade script audit (Polish Pro, v7.1), a premium design pass (v8.3), a fully-migrated Postgres backend (v9), a lip-sync delivery pipeline + template market + cost observability (v10), and a live API health board — every screen below is a real capture of the running app.

_{v10 landing — looping cinematic hero, 8 collaborating agents · 7 media engines · 3 consistency guards.}

🎙️ v10 — Lip-sync delivery · Template market · Cost guard (Stage 16)

• 🎙️ Voice & lip-sync, end-to-end — per-character voice routing (auto by name + manual pick/audition) → TTS → viseme keyframe track → measured mouth-vs-audio alignment score (Web-Audio) → drift auto-correct → pluggable engine render (wav2lip / SadTalker / MuseTalk, BYO LIPSYNC_API_URL) → written back into the timeline. One-click whole-film lip-sync with a Vision QC self-heal loop (weak shots auto re-render).
• 🧩 Template market — turn a hit project into a reusable template (style + multi-ref elements + pacing + voices), with a preview clip, ★ ratings & ♥ favorites, quality score, and one-click remix that prefills a new project.
• 💴 Cost observability + budget guard — per-project cost attribution by stage (LLM / image / video / TTS / lip-sync), saving hints, and an ok/warn/over budget guard. The publish-readiness gate is now four-dimensional: picture-vs-script · consistency · lip-sync alignability · measured mouth-vs-audio.

创作工坊 — one line → film: style presets · multi-ref element shelf · character lock · live preview · gold-neon genre gallery	模板市场 — preview clips · ★ ratings / ♥ favorites · quality score · one-click remix (carries voices)
成片质检 + 配音口型 — four-dimension publish gate · consistency trend · viseme track (animated mouth + measured alignment) · per-shot Vision scores	技术监看 · 成本归因 — per-stage cost share (video / image / LLM / TTS / lip-sync) + saving hints

Every screen above is a real capture of the running v10 app. Design language (v8.3 Taste pass — Plus Jakarta Sans + Phosphor icons, gold machined-bezel cards, spring motion, AI gold-neon genre emblems) carries across all of them.

Notes are consolidated per major version — the full minor-by-minor changelog lives in VERSIONS.md / ROADMAP.md.

📂 More modules — earlier (v6–v8) captures

Features that still ship. Director console · novel→season · finished-film station · team workspace · Cinema timeline are refreshed to v10 (live demo data). The style gallery, API health board, Polish Pro audit, and character turnaround are kept as earlier (v6–v8) captures because they show fuller sample output (the full style grid / live balances / a complete Pro audit / a 3-view turnaround sheet).

🎬 Director Console — the whole film as one control room (v6.4)

Every stage at a glance — what’s ready, what’s gone stale because you changed something upstream, and a one-click rerun that knows exactly which downstream stages it invalidates.

📖 Novel → season, with real voiceover (v6.2)

Paste a full novel; Wind Comic splits it into episodes by chapter markers (or by target length), picks a narration mode, and can render a real narration track + burnable subtitles for the whole season in parallel.

🎨 Style Gallery — 60 cinematic looks, one click (v6.3)

Lock a consistent visual identity before you generate. Search, filter by category, and apply any preset straight into the creation workshop.

🩺 API Health Board — never get surprised by a dead key (v6.7)

Live status for every model and gateway: 正常 / 额度用尽 / 配置缺失 / 不可达, with real balance read-out and a “去充值 / 补配置” hint. Keys are never stored or returned.

🩺 Polish Studio — Pro industry audit (v7.1)

Paste a draft, hit Pro: deepseek-v4-pro returns a polished script plus a full industry diagnostic — AIGC-pipeline readiness score (e.g. 85/100), style profile, first-3-second hook strength, Save-the-Cat 3-act breakdown with missing beats called out, on-the-nose dialogue lines flagged, and per-character Cameo/Seedance identity anchors so every shot stays on-model.

👤 Character Studio + Cameo IP turnaround (v6.0 / v7.x)

Every character gets a real 3-view turnaround sheet (front / three-quarter / back) with a locked structured “DNA prompt” — face geometry, skin tone, signature props, color palette, silhouette identity, full body pose — so the same actor reads identically across all 6 shots. The Cameo IP economy lets the same character travel between projects.

🎬 Finished film + 11-tab director station (v8.0)

One project, eleven tabs of cockpit-grade control: 导演台 · 剧本 · 角色 · 场景 · 分镜 · 连贯性 · 视频 · 镜头工坊 · Cinema 时间线 · 节奏分析 · 成片质检 · 技术监看 · 参数联动 · 评论协作 · 完整播放. The finished film plays right in the workspace with a 90/100 audit badge and one-click mp4 / platform export.

👥 Team Workspace (v6.5) · 🎞️ Cinema Timeline + narration track (v6.2.4)

Credit pool + per-member allocations, RBAC, real invite links.

Multi-track timeline; narration audio + subtitles burned in.

🎯 Who is this for?

🚀 Highlights · The features competitors don’t have

1. Multi-agent pipeline, not one black-box model

Director plans the story → Writer drafts dialogue under McKee structure → Style Bible Frame locks the look → Character Designer extracts an 8-dimension DNA signature of each character → Storyboard renders with Vision Audit (auto-regen on <70 score) → Video producer races multiple engines (Minimax / Veo / Kling) → Editor cuts j/l-cut on emotional beats and burns CJK subtitles.

2. The visual coherence trick — Style Bible Frame (v2.20)

We render one canonical “key art” frame from the Director’s plan, then pass it as the first --sref of every subsequent storyboard render. Net effect: all 6 shots feel like they came from the same show, not 6 random Midjourney runs. (Most competitors only carry a 2-frame rolling chain — shot 6 doesn’t know what shot 1 looked like.)

3. 9:16 by default + 12 short-drama trope templates (v2.20 P0.2)

Writer prompt detects 短剧/漫剧 genres and switches to vertical canvas + injects proven hook patterns (重生回到 N 年前 · 当街掌掴 + 秘密身份 · 系统提示音突响 · etc). McKee 3-act still backs it; tropes are the surface.

4. Real CJK subtitle burning (v2.22)

The garbled-Chinese-text-in-AI-video problem solved properly: we strip dialogue text from the video prompt (so the model doesn’t try to draw garbled glyphs) + add aggressive negatives (--no text --no chinese --no captions) + post-bake real subtitles with ffmpeg subtitles filter using a system CJK font (PingFang / Noto Sans CJK).

5. Character consistency = cref + sref + 8-dim DNA + Cameo Vision Retry (v2.21 P1.2)

Beyond reference image hacks, we run each character’s turnaround sheet through Vision LLM to extract structured features (eye shape / jaw angle / hair style / signature outfit etc.), then inject as natural-language anchor into every shot prompt. Combined with cameo-vision-retry: if a shot’s character match scores <75, we auto-regen with boosted cw.

6. Logic-Pro-style multi-track timeline with real-time collab (v3.1.1–v3.1.3)

• 3 tracks: shots / BGM / subtitle
• Real BGM waveform decoded via Web Audio API (not procedural)
• Drag-to-retime + edge handles to resize duration
• Auto-snap to neighbors within 0.4s threshold + hard-clamp on overlap
• Real-time multiplayer: Yjs awareness paints other users’ cursors live, presence avatars show which tab each collaborator is in, Y.Map locks prevent two people editing the same segment
• Project invites with viewer/commenter/editor role gating

7. Lipsync that actually works (zh / en; needs a public video URL + audio ≥ 2s)

Kling lip-sync API for talking heads, with Sync.so and Hailuo as auto-fallback. The pipeline strips dialogue from the prompt so the model only generates lip motion; we then sync the lips to the TTS audio in post. Real-machine limits: dialogue audio must be ≥ 2 seconds and the source video must be reachable at a public URL; ja/ko/ru currently degrade to no lip-sync (honest skip, surfaced in the engine-weather panel).

8. Conflict / reversal / cliffhanger pacing audit — v2 (v2.21 P1.1 → v12.275)

After Writer finishes, we score each shot 0-10 on a Chinese-conflict-word dictionary + detect emotional polarity reversals + cliffhanger keywords.

v12.275 turned scoring into diagnosis. An average score hides the difference between a story that builds and one that peaks in shot 1 — [1,2,4,9] and [9,4,2,1] both average 4. So v2 adds four things the average cannot see:

• Conflict-curve shape — least-squares slope + peak position + peak prominence, classified as escalating / flat / front-loaded / no-climax.
• Drag-segment localisation — names the exact shot range (e.g. “shots 3–5”) where conflict flatlines, instead of reporting one global number.
• Opening density — the first third is audited separately, because completion rate is decided there.
• Duration rhythm — v1 never looked at duration; v2 flags both uniform-length monotony and long-take pile-ups.

Every finding points at the shots to change. All pure functions over existing fields — no extra LLM calls, no added cost.

9. Bring Your Own LLM (v3.1.3)

Every text-LLM call (Director / Writer / Vision / Audit) goes through one OpenAI-compatible chat/completions endpoint. Want to swap to DeepSeek-r1 / GPT-4o / Claude (via OpenRouter) / Qwen-Max / local Ollama? Edit 3 lines in .env. Zero code change. See docs/llm-providers.md for the full matrix.

10. 6918 tests, TypeScript strict, no fake "coming soon"s

Every feature listed above is in main, type-checked, unit-tested, and visible at /projects/[id] if you npm install && npm run dev right now.

🥊 vs. competitors

Lineup verified 2026-10-15 (research agents + independent checkers, then I re-read both Artificial Analysis boards live and pulled every open-source star count from the GitHub API. Ledger: docs/competitive/claims.json; full analysis: docs/COMPETITIVE-GAP-2026-09.md). ⚠️ Both arenas were re-versioned twice in three weeks — text-to-video went to v2.0 and then v2.1 (Kling 3.0 1080p Pro fixed at 1000), image-to-video from v1.0 to v1.1 — so figures in earlier versions of this README cannot be compared with these. T2V with audio (v2.1): Wan 3.0 1173 > Utopai X 1166 (new; no API) > Seedance 2.5 1159 > MiniMax H3 768p 1158 > MiniMax H3 Max 1156 > FLUX 3 1145; Gemini Omni Flash 1.1 is #8 (1132) and Veo 3.1 #19 (968). T2V without audio: Wan 3.0 1134 > MiniMax H3 768p 1121 > MiniMax H3 Max 1117. I2V with audio (v1.1): MiniMax H3 768p 1141 > Seedance 2.5 1139 > Gemini Omni Flash 1.1 1134 > Wan 3.0 1127 > MiniMax H3 Max 1124; without audio: Gemini Omni Flash 1368 > Bach 1.0 Pro 1362 = Wan 3.0 1362 > MiniMax H3 1356. Vidu Q4 Preview (launched 10-07) entered the old v1.0 I2V board at #3 (1179) on 10-09 but is not listed on v1.1, even with research previews shown. Kling 4.0 is announced (up to 30s / 4K / 15 references) but not callable yet: the full model is set for October, the Flash variant has been in limited early access since 9-28, and no API pricing is published — so the column below stays on Kling 3.0, the version you can actually call (native 4K / 60fps / 15s / 6 coherent shots / built-in multilingual dialogue + lip-sync, $0.084–0.112/s as of the 09-22 check). xAI Grok Imagine Video 1.5 (Jun 16) still offers 7 simultaneous visual reference anchors (character + scene + prop + style) plus native audio and voice cloning, $0.08 (480p) / $0.14 (720p) / $0.25/s (1080p) — more reference slots than our current cref+sref pair. Veo 3.1 (first shipped 2025-10-15; 4K and vertical added 2026-01-13) can reach 60s+ via Scene Extension — but that is stitching; the single-generation ceiling is still 8–10s. It uses its own joint audio-visual generation (Lyria 3 is a separate Google music model, not Veo’s audio component), from $0.40/s. ✅ The Sora 2 API shut down on 2026-09-24 as announced (no extension) — and this project has been guarded since v12.173/207: warn before the date, auto-drop from the model chain after it (falling back to veo/kling), throw only if the chain empties (tests/v12-452-sora-sunset-runtime.test.ts runs the branch against a fake clock; tests/v12-173-sora-sunset.test.ts keeps the defaults clean). Not an open action item. 🇨🇳 The domestic camp broke through on both shot length and reference count — and two engines we already ship are behind. Seedance 2.5 (7-31) and Wan 3.0 (8-24 GA) both do native 30s single takes; Seedance accepts 50 reference inputs (token-billed, ¥42–70/M tokens), Wan 3.0 runs ¥0.42/s at 720P and converts documents straight to video (PPT/Word/PDF → 30s). Vidu Q4 Preview (10-07) succeeds Q3 as ShengShu’s flagship (briefly #3 on the old I2V-with-audio board, not listed after the v1.1 re-version): up to 15 image + 3 audio references, 2K/4K output at 10-bit, 1–16s, aimed at independent creators, small studios and production teams. ⚠️ Our own Vidu integration is still pinned to the viduq3 family (services/vidu.service.ts) — Q4 is not wired in yet. PixVerse C1 (4-08) is the most narrowly short-drama-vertical of all (storyboard-grid output + multi-speaker lip-sync). ⚠️ One finding about our own stack, which matters more than any competitor: ① both of the previous round’s self-findings are now fixed — the default video model moved to H3 in v12.402 and the two disagreeing Vidu call paths were unified in v12.403. What replaced them is a billing wall: H3 is pay-as-you-go only — MiniMax’s own pricing page says Token Plan does not cover it (re-confirmed 2026-09-22, $0.08/s at 768P, $0.13/s at 2K). On a Token Plan key every H3 call is refused, so the pipeline falls back to Hailuo-2.3 (v12.446 made that fallback actually fire — it had been dead for 16 days because it only recognised the English error text), and reference-video motion transfer (v12.448) cannot run at all. The health page now says so out loud instead of showing MiniMax all-green off a TTS probe (v12.452). ② The Sora 2 API retired on 2026-09-24 (re-confirmed on OpenAI’s deprecations page 2026-09-22, no extension). The guard has existed since v12.173/207 — warn before the date, drop Sora from the model chain after it, throw only if the chain empties — but until v12.452 the two tests “locking” it only grepped the source, so the branch had never actually executed. It is now run against a fake clock with mocked upstreams, which also exposed a gap: if the vendor stops serving before our UTC-midnight cutoff, a Sora error used to abort the whole chain instead of falling through to veo/kling.

_{How this round was checked, and what it overturned: an independent checker caught one research agent’s Elo figures off by 1–2 points; my own GitHub API pull overturned three agents’ star counts (ViMax put at ~2.7k★ — actually 12.7k★; ArcReel at 1.3–2.3k★ — 5.5k★; DramaClaw at ~1.5k★ — 6.3k★); Grok Imagine Video 1.5 Lite’s release date is contradictory (the board lists Aug 2026, trackers say Oct 1 or Oct 8), so we state none; and Vidu Q4’s “$0.014/s” launch price comes from a single article while the board listed $5.04/min at 1080p, so neither is used as a table figure. Both boards changed version between 10-09 and 10-14, so every Elo figure above was re-read on 10-14 and again on release day, 10-15 (scores unchanged; SkyReels V4 left the text-to-video board, so the models below it each moved up one place). The previous round’s two self-corrections (Veo 3.1’s dates and Scene Extension; Runway Gen-4.5 has had native audio since 2025-12-11) still stand. Elo figures are arena snapshots and move with voting.} ⭐ BYO 架构再次接住这波:榜上模型基本都开放 API,填 key 即成为本管线可调度的引擎 —— 竞品越强,本管线越强。

🔴 We overclaimed last round — four “only we have this” claims retracted (verified 2026-09-03). This is not competitors catching up; the claims were too broad to begin with: ① “EDL/AAF export is unique to us” — Descript has had it: Timeline Export covers Premiere XML, FCPXML, Reaper (EDL) and Pro Tools/Logic (AAF). ② “the only MIT, self-hostable end-to-end AI short-drama platform” — false: EvoLinkAI/ai-short-drama (MIT + Docker + novel→script→storyboard→image→video→voice→film) qualifies too, alongside LocalMiniDrama, Toonflow, huobao-drama and others. ③ “BYO multi-provider registry is unique” — PopShort.AI already ships cross-vendor models (Veo 3.1 / Kling / Seedance / Nano Banana Pro). ④ “no competitor does pacing/reversal auditing” — Volcano Drama publicly claims a multi-agent verification mechanism with 200+ shot strategies optimising conflict, reversal and climax. What survives, stated narrowly: Volcano Drama optimises inline during generation, and Descript exports but has neither pacing audit nor an open licence — so as of 2026-09-03, no single product has all three of「pacing audit hardened into an independent blocking engineering gate + EDL/AAF export + open-source self-hostable for commercial use」. Note this is narrower than last round: “pacing audit” alone no longer holds; it must be “an independent blocking gate”. ⚠️ Self-correction (2026-09-25): until v12.455 this did not match the code — the pacing audit only warned (the comment in writer-agent literally says non-blocking). v12.455 adds a real block: with PACING_GATE=block, a script that fails the audit stops before character design, is never auto-released on a timeout, and cannot be bypassed by resuming; the pacingGate request field can only tighten the server setting, and proceeding anyway takes an explicit pacingOverride: true, which is recorded on the script asset; the default is still warn-only. Also, the “EDL export” leg of the triple is no longer rare in open source: openframe (AGPL) already exports FCPXML/EDL, and ArcReel (AGPL, 5.5k★) has committed to an editing-timeline model. Open-source scale (GitHub API, 2026-10-15): OpenMontage 68.3k★ (AGPL), ViMax 12.7k★ (MIT), DramaClaw 6.3k★ (Elastic 2.0, not FOSS), ArcReel 5.5k★ (AGPL), BigBanana 2.3k★ (non-commercial), Novella AI 125★ (MIT). OpenMontage is still the one to watch, and still has no timeline-file export (EDL/FCPXML/AAF) — its “edit decision list” is an internal step before Remotion renders. Full analysis: docs/COMPETITIVE-GAP-2026-09.md. v12.214→244 双线推进。产品层:GPT Image / Nano Banana(Gemini)接入插件式图像 provider 链(issue #11,社区 @flobo3 提议,OPENAI_IMAGE_ENABLED / GEMINI_API_KEY 门控、原生 i2i 接角色一致性契约);多集连续生成补上「剧情记忆」(第 N 集 Writer 注入前几集前情提要 + 承接纪律,对标红果/阅文的 60~100 集连续,此前各集独立成篇)。平台/工程层:六轮独立对抗复检把安全洞从 CRITICAL 到 LOW 全清(SSRF 逐跳重验重定向 + IPv6 全隧道变体 / serve-file 签名能力 URL / WebSocket 鉴权 / 预算护栏),并把反复踩的「改了守卫却没跟到消费方」这个病固化成 CI 入库门禁(npm run gate:consumer,零容忍,上线即抓到 2 个人肉复检漏掉的真 SSRF)。 结论不变:生成层已是红海(竞品在出片/多镜/音频都第一梯队),Wind Comic 护城河收窄到「制作/平台层」——节奏审计、智能剪辑、字幕烧入、协作、自托管、开源、BYO。

Cells marked ⚠️ = the feature exists but in a limited / locked-down form (e.g. “you can only do this on a paid Pro tier through a UI panel”).

💰 市场与商业价值(核验 2026-08-07)

本节数据全部来自联网检索(艾媒、DataEye、克劳锐、前瞻、新华网、广电总局公开文件等),含对本项目不利的事实。我们认为写清楚赛道的真实经济结构,比堆一个漂亮的 TAM 更有用。

赛道体量

降本:本管线对标的传统成本

综合降本 80–90%。 极端案例《安徽小木匠》制作成本 2900 元、收益约 50 万元(ROI ≈ 170 倍)。

⚠️ 但必须说清楚的三件事(这决定了本项目该卖什么)

① 制作只占总成本 7.5%,投流占 70–85%。 前瞻网口径:平台广告投放 82.5%、制作 7.5%、演员 5%、剧本 1.5%。这意味着「把制作成本再砍一半」对项目盈亏几乎无影响 —— 任何以「更便宜地出片」为唯一卖点的工具,价值天花板极低。

② 单位经济已经很薄。 CPM 从 2025 下半年约 60 元/千次跌到 2026 年 15–30 元;标准 ROI 仅 1.03–1.07(东南亚出海较优,1.2–1.5)。约 90% 的 AI 短剧公司处于亏损,AI 漫剧爆款率不足 0.1%,头部 5% 项目吞掉 70% 行业利润。

③ 平台正在给「纯 AI 生成」降权。 2026 年抖音/爱奇艺/腾讯视频推出分级分账新政,削减无真人出镜的全 AI 短剧保底资源,转而扶持「真人 + AI 协作」。

于是,本项目的价值主张必须是这三条(而不是「省钱」)

1. 拉高成品率与质量下限,去够那 <0.1% 的爆款率。 节奏/冲突审计、逐镜 Vision 质检与自动重生、角色一致性(DNA + 嵌入余弦)、发布预检 —— 都是在减少废片,而废片率才是这门生意的真实杀手。
2. 能交付进专业剪辑线,支撑平台正在扶持的「真人+AI 协作」。 EDL / FCP7 XML(Premiere、达芬奇可导入)/ 真二进制 AAF(MS-CFB,面向 Avid,尚未在 Avid 实机验证)导出 —— AI 出的素材可以进人类剪辑师的工程,而不是只能整片直出。Seedance/即梦、可灵、TapNow、OiiOii 都没有结构化时间线导出(2026-09-25 核查;即梦的「去剪映编辑」送的是散素材,不是时间线)。但这不是「全部竞品的空白」:Descript 早有,开源的 openframe 也已能导出 FCPXML/EDL。
3. 自托管 + 开源 + BYO key,服务 B 端代工与品牌定制。 行业正从 C 端投流转向文旅/品牌定制 B 单(客单价数十万至百万级),这类客户对数据不出私域与成本可控有硬要求 —— 闭源 SaaS 天然做不到。

🚨 合规:2026-09-01 起的硬门槛(距今不足 1 个月)

广电总局《微短剧发展管理办法》2026 年 9 月 1 日施行:

• AI 生成/辅助制作的微短剧,须在每集显著位置标注「AI创作」或「AI辅助制作」,字体不得模糊缩小;
• 视频元数据须嵌入 AI 生成溯源信息;
• 三级分类备案 —— 一/二级须播出前取得省级及以上审核许可,无备案不得发行、不得参评。

本项目当前状态(诚实披露):该办法约束的是内容播出/发行方,不是开源工具本身;本项目不运营内容,故不直接适用。工具侧已提供的手段:v12.222 的抖音直发强标 aigc_info + 出海打包结构化 AI 声明(未确认即 422)+ 成片角标(AI_WATERMARK=1 开启,文本可用 AI_WATERMARK_TEXT 覆盖为「AI创作」)。尚未提供:视频元数据溯源字段写入。若你要把成片投放到中国大陆平台,请自行开启角标并按平台要求完成备案与标识 —— 合规责任在运营者一侧。

🆕 New in v11 → v12 — pull-replicate, smart editing & management

Real puppeteer captures of the running app (node scripts/capture-v12.mjs). 核心迭代:

• 阶段十九 AI 拉片复刻(v11.1.x):拉片五栏真值表(出厂参数,非 AI 看图猜)· 外部视频 ffmpeg 拆条 + BYO Vision 打标 · 替换工作台(「全员换猫」级全局替换)+ 复刻起片(保原片镜头结构/时长)· 复刻保真度对照 · 存为私有模板。
• 项目/资产管理(v11.2.0):我的项目 / 我的资产 删除 + 下架(级联清理 + 属主守卫)。
• 模型雷达(v10.6.3):一键扫描各 API 最新模型 + 同家族自动升级(四护栏 + 回滚)。
• 钩子审计三指标(v10.6.2):开场 3 秒钩子 / 集尾悬念 / BGM 卡点对齐率。
• 阶段二十 A 智能剪辑(v12.0.x,五刀全交付):卡点剪辑(切点吸附音乐拍点)· 情绪节奏曲线(峰值镜 breathe / 动作镜快切)· 侧重强调(关键镜不压 + 沉稳转场)· 转场审美(按镜头关系选转场)· 一句指令调风格(v12.0.4,BYO)——「快节奏燃向」/「慢叙抒情」一句话调 pacing 力度 + 转场软硬,无 key 走规则、配 key LLM 解析自由文本。
• 阶段二十 B 预览音频(v12.1.x):片段预览叠播配音(静音裸片 + 同步配音轨)+ 成片音频体检自愈(ffprobe 缺流补轨)。
• 阶段二十一 角色/资产一致性升级 · 全局资产记忆库 v2(v12.2.x,五刀全交付):对标同构竞品 OiiOii「角色高维特征向量 + 跨场景一致性」——名称归一修 DNA 漏注入 · DNA/场景锚落库(rerun/重启复用、早镜不漏) · 给 global_assets.embedding 死列通电(BYO 文本嵌入 + 内存余弦检索) · 建角色入口「相似角色」推荐一键复用(防重复建/跨集漂移) · 身份漂移检测(逐镜视觉 embedding 余弦距离标 outlier 漂移镜)。全程无 key 走确定性地板(精确名+文本匹配),有 key 向量增强,诚实降级。
• 安全加固(v11.1.4):JWT 公开兜底密钥根除(进程级随机密钥,旧泄露值作废)。

我的项目 · 删除/下架管理	拉片表 + 复刻工作台

素材库 · 资产管理	API 健康 · 模型雷达

钩子审计三指标	一句指令调剪辑风格(v12.0.4)

片段预览音频:三态徽章 + 带声试听(v12.1.2)
	左:带配音(TTS 叠层)+ 带声试听开关 · 右:片段无独立音轨(成片含配乐+配音)

完整逐版本核心迭代见 VERSIONS.md;阶段计划见 docs/stage18–docs/stage20。

🎬 Screenshots

📸 v12.425 资产真机截图(captured 2026-09-06)

全部由 node scripts/capture-v12-425.mjs 对着本机真实数据跑出来。脚本带空壳检测与近重复检测: 页面没渲染出来、或者「点了 tab 但视图没变」的,一律删掉并如实报告 —— 一张截图冒充两张,比少一张更糟。

📚 完整 18 张 + 每张的实现逻辑与工作流架构说明:docs/SCREENSHOTS-v12.425.md (单独成文是为了不让首页替访客扛下载量 —— README 引用媒体有 12MB 预算门禁。) v12.426 补记:上面「我的项目」那张(docs/screenshots/v12/01-my-projects-manage.jpg)已重拍。 旧图里卡片大半是渐变占位 —— 不是没作品,是封面被冻结成了一次失败的快照:流水线末尾抄一份第 1 镜的 imageUrl 就再不重算,于是出图全挂时的 mock 兜底图被永久写死(30 个项目里 12 个如此,其中 3 个明明各有 11~12 张真分镜还活着)。现改为读时解析,并清掉了 21 个夹具/无素材项目(下架,可一键恢复)。

角色转身图 · v12.425 修的就是这里	导演台 · 全链路控片
立绘原生 896×1152 竖构图,修前被塞进 355×200 横框 + object-cover,裁掉 56%,必须点全屏才看得到完整图。现在框比例跟素材走、填充改 object-contain。	四个环节各自可寻址:剧本 / 角色·场景 / 分镜 / 成片,任一环单独「编辑」或「重跑」,并预告重跑的下游影响。

拉片分析 · 五栏出厂真值	分镜规格 · 出厂参数 + 一致性仪表
每镜给出叙事要素 / 时间 / 镜头语言 / 影像处理 / 声音五栏,且是流水线生成时的真实摄影语言,不是 AI 事后看图反推;可导出 CSV / 剧本册 MD / PDF,也能回灌外部片子做复刻。	画幅取项目创建时选定的 projects.aspect(整片生成、分镜构图、导演台同一来源),格式条只读显示;成片帧率按格式条设置(24 / 25 / 30fps),EDL / AAF 导出的时间码也用它;片段重拍按该镜原片的真实帧率缝合;色彩空间写进之后出的分镜图提示词(默认「不指定」,不加任何色彩描述);安全框开关项目页的竖屏安全区预览叠层(只影响预览,不进成片)。右侧逐镜一致性打分,低分镜一键跳转重生。

📸 v12.402 – v12.416 screenshots (captured 2026-09-03)

All captured by node scripts/capture-v12-416.mjs against real local data (1440×900 @2x, downscaled to 1600px wide). The script has blank-page detection and near-duplicate detection: a screen that failed to render, or one where “the tab was clicked but the view never changed”, is deleted and reported honestly — one screenshot masquerading as two is worse than one missing.

Only one project-page shot: clicking the tabs and scrolling the body both leave the view unchanged. This conclusion was wrong and is retracted (v12.425.) The tab bar sits below the fold, so a coordinate click missed it; scrollIntoView followed by el.click() switches tabs fine. See the v12.425 section above for 12 shots taken that way.

Measured performance (local dev server, curl time-to-first-byte): landing 46 ms; console pages (projects / assets / engine health / usage) 22–38 ms. Tests: 5294 passing · 0 failing; preflight 10/10; CI all five jobs green.

Below is the foundational v3 pipeline (the v6 studio screens are in the New in v6 section above). Every panel is a real puppeteer capture of the running app (run node scripts/capture-screenshots.mjs / node scripts/capture-v6.mjs to refresh).

Workspace overview

The 创作总览 dashboard: 99 projects + 4 case studies + recent activity feed + system status (engines in use, model versions).

Asset library

Cross-project reusable: 角色 / 场景 / 视频 / 音乐 / 字幕 / 模板 — 1467 assets in this demo project.

Project library

Every short film with auto-generated cinematic covers + status badges + quality donut.

Creation workspace — live multi-agent canvas

The whole pipeline as a live agent flow: Writer / Character Designer / Scene Designer / Storyboard Artist / Video Producer / Editor nodes wired together with progress streaming per node, plus a chat side-rail showing every agent message in real time.

Per-project script + shot list with beats

The 剧本 tab: every shot with duration, emotion tag (警觉 / 凝重 / 惊恐 / 暴风的沉着 / 镇定的专注…) and a one-line beat note (从表面到深层警觉 / 从无知到悉知威胁 / 从警戒到遭受袭击 …) so the rhythm of the cut is legible at a glance.

🆕 Cinema Timeline (v3.1.1–v3.1.3 — multi-track + collab)

3-track layout (SHOTS / BGM / SUBTITLE), drag-to-retime, double-click subtitles to rewrite, drag edges to resize, real BGM waveform (Web Audio decode), live other-user cursors with name labels, segment lock indicators.

🆕 Pacing Analysis (v2.21 P1.4)

KPI: 平均冲突分 / 反转数 / 通过状态. Per-shot conflict-score bar chart with reversal arrows + emotional polarity icons. Color-coded green (≥7) / amber (4-6) / red (<4). Below: actionable warnings + suggestions.

🆕 Comments + @mentions (v3.0 P0.1)

Project-level + per-shot threaded

README 内容过长，已截取前 64 KB。完整内容请前往 GitHub 仓库查看。