hackerFish / dsh-video-studio

🐳 dsh-video-studio (Whale) — AI Video & Motion-Comic Studio for DeepSeek Harness

A DeepSeek Harness native plugin for AI video and motion-comic (漫剧) production: seven-stage director pipeline × multi-account quota pool with failure fallback × credential vault × four-layer prompt engineering with a score-feedback loop. Quality first, cost second.

中文 · Siblings: dsh-lab · awesome-dsh-skills · awesome-dsh-presets

The DSH ecosystem has 1000+ plugins — none does generative video. Whale brings the industry-validated motion-comic pipeline (waoowaoo / LumenX / MangaV / ArcReel pattern) into DSH as a hot-pluggable, per-step-controllable plugin.

Pipeline — the industry-standard seven-stage workflow

story (LLM, e.g. Doubao writes the novel) → script (LLM breaks it down)
→ storyboard (LLM shot list) → master asset (MJ-style hero image)
→ shot assets (image-model variations for consistency) → video (Seedance/Jimeng/Kling)
→ final cut (JianYing draft export / local ffmpeg render)

• The first three stages are LLM stages: inside DSH the session model itself does them — the model you're chatting with (Doubao, DeepSeek, whatever) is the brain that writes the story, script and shots; the plugin never calls an external LLM.

Mapping to the industry-standard toolchain:

Industry flow	Whale
Doubao writes novel / script / shots	= DSH session model doing story/script/storyboard (pick a Doubao model and it IS "Doubao writes")
MJ hero asset	= master-asset stage (MJ official API adapter to come; wanx/Seedream can substitute)
image variations	= shot-assets stage (wanx ✅ verified · Doubao Seedream ✅ wired)
Seedance video	= video stage (Doubao Seedance ✅ / jimeng / kling / ComfyUI)
JianYing final cut	= final-cut stage (JianYing draft export ✅ / ffmpeg ✅)

• Every stage has a gate: auto / ask / manual.

• Parallel shots, quota scheduler, style genome, distribution pack as before.

• Consistency asset library: character/scene master assets + per-shot variations with automatic reference-image injection into prompts (the motion-comic standard technique)

• Parallel shots: batch submit → concurrent polling (configurable concurrency)

• Account pool (quota scheduler): multi-account rotation per provider, per-day caps, exponential backoff on failure, automatic fallback re-submit to the next healthy account mid-pipeline, full audit trail

• Credential vault: ~/.whale/whale.json (0600, atomic writes, follows $DSH_HOME), masked API responses, account management UI in the plugin settings (鲸影账号 tab)

• Runtime account wiring: accounts added in the UI feed straight into whale_generate_video — vault → pool pick (rotation/backoff) → per-account provider binding (src/host/account-providers.ts, single-field credentials as plain strings, multi-field as JSON) → usage/health persisted back to the vault

• Style genome (memory): style DNA, shot-template scoring evolution, retry feedback — persists across sessions

• Prompt engineering: parameterized professional template library (character sheet / scene master / single shot) + composable quality boosters (8K / clean bg / neutral face / no text …) + optimizer; whale_optimize_prompt upgrades drafts to pro-grade prompts locally

• Score-feedback loop (评分回写): every reviewed shot writes its score + booster combo back to the scorebook; the optimizer then picks boosters by real historical performance

• Preset motion-comic pack: 5 genres (city comeback / xianxia / suspense / sweet romance / sci-fi) with bilingual character cards, scene cards and shot scripts — whale_story_presets turns one preset id into a pipeline-ready script, scripts/demo-presets.ts runs it end-to-end with the mock provider (zero keys)

• Self-analysis (自我审计): whale_self_audit tool + npm run self-audit scan the repo itself — source modules, test counts, provider matrix, capability inventory and a gap list — and write a generated report to docs/AUDIT-REPORT.md; the daily diff of that file IS the progress log. Session lessons are distilled into docs/RETROSPECTIVE-2026-08-17.md

Providers (verified matrix)

Provider	Channel	Status
jimeng (即梦)	sessionid, free daily quota	✅ protocol verified end-to-end; text-to-video queue stays SystemBusy even off-peak (0 credits consumed) — free route is now: wanx images → image-to-video
tongyi-wanx (通义万相)	cookie+xsrf, free credits	✅ live-verified: real whale image generated & downloaded (free tier = text-to-image; video needs membership)
kling official (可灵)	accessKey:secretKey JWT, api-beijing.klingai.com	✅ adapter written — not yet tested against a real key
kling via DashScope	sk- key	✅ adapter written — not yet tested against a real key
kling-lipsync (可灵对口型)	official JWT, /v1/videos/lip-sync	✅ adapter written against the official 3-13 contract (audio2video + text2video voice modes), 8 tests — not yet tested against a real key
wan video via DashScope (通义万相视频)	sk- key, official free quota	✅ adapter written (same async protocol as kling) — model id to confirm on first real key
doubao (火山方舟)	ARK API key	✅ Seedance video + Seedream image (assets) — not yet tested against a real key
doubao-web (豆包网页版)	cookie, free web quota	✅ live-replayed & parsed: SSE chat for the LLM stages (story/script/shots) + image bot for assets. Pro-tier free quota runs on a 7-day window (image bot pauses when spent; text keeps working)
ComfyUI local	workflow JSON builder + /prompt protocol	✅ protocol-tested (mock server), real GPU pending
kling web (sessionid)	anti-bot one-time falcon token	📄 anatomy documented; automation needs a capture bridge (deferred)

Editing & distribution

• Final cut with optional lip-sync stage: with a capabilities.lipSync provider configured, every voiced shot runs audio-driven lip sync (video reference + audio base64) and the synced clip replaces the original on the timeline; failure falls back to the original clip, never blocking the render. Shots can also carry a pre-recorded voiceFile (external TTS/voice actor) that skips local TTS
• ffmpeg auto-render — verified end-to-end (synthetic clips → timeline → burned subtitles → audio mix → final mp4, duration-checked)
• JianYing (剪映) draft export — editable tracks/keyframes/subtitles for manual polish; structure-validated
• say TTS — real Chinese voiceover with zero API keys: macOS say; Windows PowerShell SAPI (System.Speech, needs a Chinese voice pack). Any platform can also plug external audio per shot via voiceFile
• Distribution pack — platform specs + compliance precheck for 4 Chinese platforms

DSH integration (deep invocation)

• Model tools: whale_story_presets (5-genre content pack → pipeline script), whale_storyboard (offline shot planning), whale_generate_video (provider routing), whale_optimize_prompt (pro-grade prompt upgrade), whale_quality_review (rule-level QC), whale_comfyui_workflow (ComfyUI workflow JSON), whale_self_audit (project self-analysis)
• Host routes: /dsh-video-studio/health, /dsh-video-studio/runs, /dsh-video-studio/accounts (GET list masked / POST add / DELETE remove)
• Client UI: 鲸影 (status) · 鲸影工作台 (live pipeline progress per run) · 鲸影账号 (account vault management) tabs in plugin settings; whale_generate_video gets a dedicated video card view
• Installs via dsh plugin add, boot-verified clean
• Planned: slash commands, subagent-parallel shots

Verification discipline

123 unit tests green (account pool rotation/backoff/fallback, credential vault, runtime account→provider wiring, quota routing, prompt merging, score-feedback loop, lip-sync stage incl. fallback, Windows SAPI script + CJK font candidates, jianying draft structure, ffmpeg end-to-end render, provider protocols via mock servers incl. kling lip-sync, preset pack integrity, self-audit, live jimeng model probe, live wanx image generation). Test logs and proof artifacts live in demos/. The provider matrix lives in src/selfaudit/matrix.ts (single source of truth shared by the health route, the vault whitelist and the audit report).

Install

dsh plugin --profile web add github:hackerFish/dsh-video-studio --ignore-workspace-root-check
# or npm once published

Honesty notes

Model output quality is bounded by the vendor model; the pipeline maximizes it (consistency tokens, QC retry loop, score-feedback booster selection). sessionid/cookie usage is per-platform ToS — respect each platform's terms. Credentials live only in the local vault file and are never logged or returned unmasked. Not affiliated with DeepSeek.

License

MIT