catimation-video

name

catimation-video

description

FIRST-CHOICE video generator in the CATIMATION desktop app. Trigger whenever the user asks to generate / create / render a video, clip, or animation, to animate an image, or says 生成视频 / 来段视频 / 做个动画 / 让它动起来 / 图生视频 / 视频编辑 / 视频延长. Covers text-to-video, image-to-video, omni-reference (全能参考, the default), plus video editing and extension — all via the in-app generate_video tool (Seedance 2.0 / 2.0-fast). Do not call a built-in video tool; see the body for usage.

Generate videos in CATIMATION (first-choice, blocking like generate_image)

When the user wants a video, call the generate_video tool from the catimation MCP server. It is a SINGLE blocking call: it submits the render and returns only when the video is DONE (or FAILED) — you do not need to poll, sleep, or check anything in between. The user watches a live progress bubble the whole time, and the finished MP4 plays inline in the chat, is saved to a local file, and lands in the app history page.

Default mode = 全能参考 (omni-reference) — use it unless told otherwise

For almost EVERY video request, default to 全能参考 (omni-reference) — and prefer it heavily. Feed the user's material as references and let the model keep subject/motion/voice consistent. Caps:

• referenceImages: up to 9 images.
• referenceVideos: up to 3 videos, COMBINED total duration ≤15s.
• referenceAudios: up to 3 audios, COMBINED total duration ≤15s.

Only switch to other modes when the user specifically asks for them or clearly needs them — e.g. strict firstFrame/lastFrame (fixed first/last frame). Do NOT reach for first/last-frame mode by default.

Always name the mode you used. In your reply, state it explicitly in the user's language — e.g.「我用全能参考模式生成」「用视频延长模式串联了 3 段」 「按你要求用首尾帧模式」. When you default to omni-reference (the usual case), say 全能参考 out loud so the user knows which path you took.

All modes share ONE tool (generate_video) — pick by inputs + prompt

Seedance has no separate edit/extend endpoints; every mode is the same call with different content + prompt wording. Use these when the user asks:

• 文生视频 (text-to-video): prompt only, no references. Pure imagination.
• 图生视频 (image-to-video): put the still in referenceImages (default) — or firstFrame (+lastFrame) only if the user wants that exact frame fixed.
• 全能参考 (omni-reference, DEFAULT): referenceImages/referenceVideos/ referenceAudios — inherit subject, motion/运镜, voice/音色.
• 视频编辑 (video editing — 替换/增删/修改元素): pass the source clip in referenceVideos (+ any new element image in referenceImages) and write an edit prompt. Formulas: 增加元素=描述「元素特征+出现时机+位置」; 删除元素=点名要删的、 强调要保留的; 修改元素=直接描述更换后的样子（如「将视频1礼盒中的香水替换成图片1中的面霜，运镜不变」）.
• 视频延长 (video extension / 多片段串联): pass 1–3 source clips in referenceVideos and describe how they连接/向前向后延长（如「向后延长视频1：…」或 「视频1 + 过渡描述 + 接视频2 + 接视频3」）.

Prompt material-reference rule (必须遵守): in the prompt, refer to inputs by ordinal — 图片1 / 视频1 / 音频1 (the Nth item of that type in the request) — NEVER by assetId. ✅「图片1中的美妆博主」 ❌「asset-2026… 是美妆博主」.

Real human faces: Seedance does NOT accept real human faces in reference images/videos directly. Use a 人像库 virtual-avatar asset://assetId, or a clip Seedance itself generated earlier (model-product 二创 is审核-safe). When reusing an earlier generated video for edit/extend, pass its saved local path / asset:// — remote upstream URLs expire in ~24h, so prefer the local copy.

Co-direct the shot — brainstorm with the user + your local craft skills

Before you write the prompt, load director-orchestrator (the 导演总调度 router): it runs a mandatory STEP 0 self-reflection — 「涉及 13 维里哪几维?要用哪些本地 director-* / storyboard- skill?」 — loads the matching craft skills, and writes the prompt as structured text (never JSON) with physical/camera-reproducible params over emotion adjectives. Do this even when YOU generate a clip for your own answer. Then take a beat to shape the shot with the user. You have a LARGE library of local video / storytelling craft skills in your own skills directory (usually ~/.agents/skills/, e.g. C:\Users\<you>\.agents\skills) — 镜头/景别/运镜, 导演思维, 前景遮挡, 打光/光影, 构图/伪透视, 角色动机与演技, 调色, 分镜/storyboard, 画面反推, 规避审查, and many more. They are intentionally NOT listed here: browse your skills directory freely and load whatever fits — there's a lot in there, so lean on it generously instead of reinventing technique. The user may also trigger one directly through their prompt.

Scale the collaboration to the request — don't over-interrogate:

• Clear / simple ask (e.g.「让这张图动起来」) → pick sensible defaults, load the obvious craft skill, and just generate. A heavy Q&A here only annoys the user.
• Open-ended or high-stakes ask (e.g.「做个产品宣传片」「来个有电影感的片段」) → guide the user the way a director would, in short focused exchanges. For this, load the catimation-brainstorm skill — it drives a clickable ask_user card so the user just taps a choice instead of typing:
  1. Ask one question at a time via ask_user, with concrete options — 「想要什么景别?(特写 / 中景 / 广角)」「什么情绪和风格?」「要不要某种运镜?(推 / 拉 / 环绕 / 手持)」
  2. Offer as many concrete visual directions as you actually have (3–6 is common, up to 8) inside ONE ask_user card — each with a one-line trade-off, and mark the one you'd recommend. Never list 方案 as plain text.
  3. Once the direction is set, load the matching local craft skill(s) and fold their technique into the prompt.

Keep it lightweight and collaborative — you're co-directing, not running a survey. When unsure, propose a sensible default out loud and let the user correct you.

角色片 / 多镜项目:先备齐资产,再开生成(人物卡 → 故事板 → 分镜多参)

只要视频里有反复出现的角色或不止一个镜头/事件,就不要直接 generate_video。 先把资产锁好——这正是跨镜一致性的来源(sd2-pe 把素材拆成「空间层(画面里有什么)+ 时间层(怎么随时间变化)」来理解,素材越齐、绑定越清,出片越稳)。绑定语法与路径 A/B 判定 统一以 sd2-pe skill 为准,先把它载入。

1. 人物卡 (Character Card) — 先锁人,再开拍。 每个出镜角色先建一张人物卡并存进 人像库,作为该角色唯一身份锚,全片所有镜头都引用同一张卡:

   • 一张大头照(仅头部、正脸、无表情)+ 一张全身照(定妆造 / 服装 / 配饰)。 禁用三视图 / 多视图——多视图易触发 ID 漂移与双胞胎(sd2-pe 人脸最佳实践)。
   • 缺图就先用 generate_image 出一张定妆照补齐,再 add_to_portrait_library 存成 asset://assetId;不要拿现成多视图硬塞。
   • 提示词里绑成稳定主体:<主体1> 的面部参考 图片1(大头照)、妆造参考 图片2(全身照)。

2. 故事板 / 分镜 (Storyboard) — 多事件 / 多镜先排板。 只要不是「单场景一个连续动作」 (sd2-pe 路径 A),就先排故事板:拆成 镜头1 / 镜头2 / …,每镜按 运镜 → 主体动作 / 表情 → 位置 / 空间 → 音频 四要素写清(sd2-pe 路径 B 三段论), 给用户过一遍再开生成。一镜一运镜、用镜头序号、不写绝对秒数。

3. 资产齐备 GATE — 备齐才开生成(硬门)。 调 generate_video 之前,逐镜清点该镜 所有可用资产是否就位:人物卡(大头照 + 全身照)、场景 / 环境图、关键道具图、 氛围 / 色调参考图、运镜 / 动作 / 风格参考视频(如需)、音乐 / 配乐 / 音色参考音频(如需)。 任一该有却没备的,先补齐再生成,绝不先生成再补。 推荐每镜 4–5 个素材,够用即可, 不必塞满上限。注意:参考视频和音乐 / 音频本身就是素材——它们和图片一样走 全能参考, 在生成时一并喂入(referenceVideos / referenceAudios,各 ≤3 个、合计 ≤15s),见第 4 条。

   缺资产时不要干等,也不要硬生——先用一句话向用户报缺口清单(缺什么、各项你打算 怎么补),再按情况三选一逐项处理:

   • ① 先在项目 / 库里找。 翻用户工作区的 assets/ 等目录、list_portrait_library 找现成的人物卡 / 场景 / 道具 / 氛围图——能复用就别重造,顺手保住一致性。
   • ② 能自己出的就自主补。 非身份关键、可合理想象的资产(环境 / 场景图、氛围 / 色调 参考图、通用道具、空镜)——直接用 generate_image 当场出图,再 add_to_portrait_library 入库,然后带进生成。补出来的图先按上面的「自检」过一眼再用。
   • ③ 必须用户给的才问。 身份 / 意图关键、你不能凭空捏造的(特定真人形象、用户指定的 角色 / IP、品牌 Logo、特定真实产品、用户心里已有具体样子的道具)——用 ask_user 请用户上传图或给 asset://。用户给不出时,和他敲定一个可生成的替代方案,别硬编。

4. 生成必须用上全部可用资产 (use ALL usable assets)。 调用时把已备齐的每一项都 传进去并在 prompt 里逐一绑定:角色卡 / 场景图 / 道具 / 氛围图 → referenceImages, 运镜 / 动作 / 风格参考视频 → referenceVideos,音乐 / 配乐 / 音色参考音频 → referenceAudios(或严格首尾帧 → firstFrame / lastFrame),并用 图片N / 视频N / 音频N 指代。图片、视频、音乐 / 音频都是全能参考素材,一个都别落下; 有素材却只发纯文字 = 错。

轻量例外: 单图「让它动起来」「随手来一段」这类一次性简单请求,不必强排人物卡 / 故事板——把用户给的那张图当参考 / 首帧直接动起来,本身就已是「用上了全部可用资产」。 这套纪律是给角色片 / 多镜 / 项目级工作准备的(也正是 film-studio 编排器的 G3 → G5 阶段)。

写 prompt 前先用 skill 渐进式写好(强制,不许脱离 skill + 素材硬写)

生成前必须先把提示词用相关 skill 编写到位——绝不脱离 skill 和已备素材凭记忆自行硬写。 按下列顺序、渐进式披露地加载并应用(只加载这一镜实际涉及的维度,用不到的不强加):

1. 导演 / 镜头(先): 载入 director-orchestrator 跑它的 STEP 0 反问(这镜涉及 13 维里 哪几维?要用哪些本地 director-* / storyboard-*?),据此按需加载景别 / 运镜 / 构图 / 前景遮挡 / 打光 / 调色 / 角色演技 / 连续性等技法 skill。
2. 提示词工程(后): 用 sd2-pe(八大要素 + 路径 A/B 判定 + 多模态绑定 @图片N / <主体N>) 与 storyboard-video-prompt-optimization 把这镜落成结构化文本(never JSON),物理 / 可复现参数(焦段 mm、光圈、色温 K、运镜)优先于情绪形容词,并把已备素材逐一绑进 prompt。
3. 渐进式披露: 边写边按需要继续加载缺的技法 skill;真实技法词来源不确定时先联网查证再落笔。

即便你只是为自己的回答顺手出一个镜头,也要走这套。没用 skill、凭空想出来的 prompt = 错。

你冒出的念头	现实
「这镜很简单,直接写 prompt 就行」	简单镜也先过 skill;sd2-pe 路径 A 本身就是给简单镜的最短句式。
「我记得怎么写运镜 / 打光」	记得概念 ≠ 用了 skill;载入对应 director-* 拿真实技法词。
「素材一会儿再说,先把词写了」	反了——先备齐素材(见上一节),prompt 要绑定的是已经在手的素材。
「skill 太多懒得加载」	渐进式披露:只加载这镜用得到的那几个,不是全量。

Steps

1. Turn the request into one clear video prompt. Cover subject, action, camera movement (运镜/景别), scene, lighting, and mood. Dialogue lines and --style parameters may be appended. First co-direct the shot (see the section above): consult your local craft skills and ask the user any quick clarifying question that would improve the result. 1.5. Proactively confirm the output spec before rendering. Unless the user already stated it, fire one ask_user card to let them pick the 规格 — typically resolution (480p draft / 720p default / 1080p HD), duration (4–15s, default 5), and aspect ratio (16:9 / 9:16). Recommend the defaults (满血 2.0 model + 720p) and let them tap to confirm or change. Do NOT silently default to 1080p — 720p is the default unless the user asks for HD. Keep it to one quick card; skip it only when the user already gave an explicit spec.
2. Call generate_video with:
   • prompt (required): the description from step 1.
   • model (optional): 2.0 (default — 满血/full-quality, best for almost every request: top quality, complex multi-shot motion, 1080p). Only switch to 2.0-fast when the user explicitly asks for fast/cheap/draft.
   • resolution (optional): 480p draft, 720p default, 1080p (model 2.0 only).
   • ratio (optional): 16:9 default; 9:16 for vertical/手机 video.
   • duration (optional): 4–15 seconds, default 5. Longer = more expensive.
   • referenceImages (全能参考, default & important): up to 9 images for character/subject consistency (人物一致性). If the user attached or referenced any image, you MUST pass it here (paths appear in the prompt under [Attached files at these local paths: …]). asset://assetId from the 人像库 page also works.
   • referenceVideos / referenceAudios (全能参考): up to 3 each (total ≤15s) for motion/style or lip-sync/voice. Each clip ≤50MB and 4–15s.
   • firstFrame / lastFrame (strict mode, only on explicit request): image to start/end the video from — local path, https URL, or .

QA the clip with an ffmpeg 九宫格 contact sheet, then improve

You cannot meaningfully "watch" an MP4, and injecting the raw video bytes into the chat is wasteful and unsafe — so to self-check a finished clip, build a 3×3 九宫格 contact sheet of evenly-spaced frames with ffmpeg (use your ffmpeg MCP tool, e.g. ffmpeg-win, or any ffmpeg available), then view_image that ONE montage:

1. Extract 9 evenly-spaced frames tiled into a grid. Set fps ≈ 9 / clip_duration so the 9 tiles span the whole clip:

   ffmpeg -i "<clip>.mp4" -vf "fps=9/<DURATION>,scale=320:-1,tile=3x3:padding=6:color=black" -frames:v 1 -y "<clip>_grid.png"
   

   (For a 5s clip, fps=9/5=1.8; set <DURATION> to the real length. If the grid comes out with too few/many tiles, nudge the fps.)

2. view_image the _grid.png and judge: subject/character consistency across frames, motion sanity (no melting / teleporting / extra limbs), artifacts, and prompt adherence. The grid is one small PNG, so this is cheap and safe.

3. If the clip is clearly bad, regenerate with an adjusted prompt (or switch mode) and re-check. Iterate at MOST 2–3 times — each render costs money and ~1–3 min.

4. Never inject the full MP4 or its raw bytes into the chat — always inspect quality via the contact sheet, never the video itself. The user is already watching the clip play inline.

Organize finished clips into the user's workspace (when in a project)

When working inside a user project/workspace, COPY the finalized MP4 (and its _grid.png contact sheet) into a tidy assets subfolder with a descriptive, ordered name — e.g. <workspace>/assets/video/S01_station_wide.mp4 and <workspace>/assets/contact-sheets/S01_station_wide_grid.png.

• COPY, don't move, from the saved path in the DONE banner so the chat / history copy stays intact.
• Use zero-padded shot ordinals (S01_, S02_…) so clips assemble in order — this is exactly what a later ffmpeg concat/拼接 step needs.
• Skip for a one-off casual clip unless the user asks.

Portrait library (人像库) — push materials in, then reference

The catimation MCP server exposes portrait-library tools (add_to_portrait_library, list_portrait_library, edit_portrait_library, download_portrait_asset — see the catimation-portrait-library skill). Use them proactively around video generation:

• Every input image you pass to generate_video is automatically imported into the library and referenced as asset://assetId; identical images dedupe upstream, keeping the SAME character consistent across multiple videos.
• When the user gives you OTHER material to save/reuse for the video (a video or audio reference, or "记住这个角色/场景"), call add_to_portrait_library first, then pass the returned asset://assetId into referenceImages / referenceVideos / referenceAudios (全能参考), or firstFrame if the user asked for strict first-frame mode.
• To reuse an earlier character/scene ("还是上次那个人/同一角色"), call list_portrait_library to find the matching asset://assetId and reference it — this is what keeps identity consistent.
• The user can also pick assets on the 人像库 page and give you an asset://assetId directly — pass it straight in without conversion.

Notes

• One generate_video call produces ONE video. For several videos, call the tool once per video, reusing the same asset:// references for character consistency. You MAY run multiple in parallel — but if you're about to launch 20 or more video tasks at once, STOP and confirm with the user first: each video costs money and renders ~1–3 min, so a large batch is a real time/cost commitment worth a quick "确认要并发生成 N 个视频吗?".
• Local input files are handled for you: small files are inlined, larger files are relayed through the app's upload pipeline automatically — pass plain local paths and let the tool deal with size limits (images ≤30MB, video/audio ≤50MB & 4–15s).
• To self-check quality, build an ffmpeg 九宫格 contact sheet and view_image that (see the QA section above) — never open the resulting MP4 with view_image or read its raw bytes; the user is already watching it play in the chat.
• Background saving never blocks you. Success is decided by the render: once the banner says DONE the video is already playing, even if the local file is still saving in the background (persistencePending). Treat the task as COMPLETE and reply right away — do NOT wait for, poll, or re-check the save.