⏳ This skill is pending AI review.
Scores will appear once the review pipeline completes.
seedance-2-5-video-director
Design, derive, optimize, diagnose, and rewrite scripts or copy-ready prompts specifically for Dreamina/即梦 Seedance 2.5. Use for new text/image/video/audio-to-video concepts, exact-timeline videos, 30-180-second Long Video, extension, video editing, Clay Renderer, seamless transitions, multi-grid storyboards, realistic acted dialogue, relationship drama, sports or variety-show live camera grammar, character and wardrobe locking, example-structure adaptation, voice/timbre reference, BGM removal,
Choose how to use this skill
You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.
1. Native installer
This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.
Do not guess an installer command or replace an existing version without reviewing the diff.
2. Complete package recommended
Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.
No complete ProSkills package is published for this listing yet.3. Prompt-only
Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.
Need only the instruction file?
Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.
No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.
// RATINGS
Not yet listed on ClawHub or SkillsMP
// README
Seedance 2.5 Video Director
面向 Dreamina/即梦 Seedance 2.5 的视频导演 Skill。它将文本创意、人物参考图、视频、音频、分镜或旧提示词整理成可执行的视频脚本、导演方案、诊断结果和复制即用的提示词。
非字节跳动、Dreamina 或即梦官方项目。本仓库不会调用付费生成接口,不会自动消耗积分。
功能
- 4–30秒普通视频与精确30秒时间轴
- 30–180秒 Long Video
- 视频前置/后置续写
- Smart Edit、标记区域编辑和上传视频编辑
- Clay Renderer 粗白模/精白模渲染指令
- 两段视频无缝转场
- 多格分镜动画
- 人物身份、体型、服装和配饰锁定
- 多人物、对白音色、BGM移除与声画同步
- 真人情感表演、情侣戏、竞技综艺直播镜头和障碍物物理
- 提示词冲突诊断与按原结构改写
安装
推荐:一行安装
已安装 Node.js 的用户,在 Windows、macOS 或 Linux 终端执行:
npx skills add liyue-aigc/seedance-2-5-video-director -g -y
这条命令会从 GitHub 获取 Skill,并安装到当前用户的 Agent Skills 目录;-g 表示全局安装,-y 表示跳过交互确认。它使用 npm 附带的 npx 启动安装器,但本仓库本身不是需要加入项目依赖的 npm 软件包,因此不使用 npm install。
如果电脑上没有 Node.js,也可以直接在 Codex 中发送:
请使用 $skill-installer,从 GitHub 安装 liyue-aigc/seedance-2-5-video-director。
手动安装(备用)
Windows PowerShell
$skillRoot = Join-Path $env:USERPROFILE '.codex\skills'
New-Item -ItemType Directory -Force -Path $skillRoot | Out-Null
git clone https://github.com/liyue-aigc/seedance-2-5-video-director.git (Join-Path $skillRoot 'seedance-2-5-video-director')
macOS / Linux
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
git clone https://github.com/liyue-aigc/seedance-2-5-video-director.git \
"${CODEX_HOME:-$HOME/.codex}/skills/seedance-2-5-video-director"
安装完成后重新打开Codex会话,并在提示词中调用:
使用 $seedance-2-5-video-director,把我的视频创意整理成可复制的 Seedance 2.5 提示词。
安装器明确提示刚完成安装时,Skill会展示一次学习指导;普通新对话会直接处理任务,不会重复教程。需要时可主动发送“显示 $seedance-2-5-video-director 使用教程”。
更新
通过 npx skills 安装的用户:
npx skills update seedance-2-5-video-director -g -y
手动克隆的用户:
git -C "${CODEX_HOME:-$HOME/.codex}/skills/seedance-2-5-video-director" pull
Windows默认安装路径更新:
git -C (Join-Path $env:USERPROFILE '.codex\skills\seedance-2-5-video-director') pull
快速使用
使用 $seedance-2-5-video-director。
任务:生成20秒、16:9的水上竞技综艺片段。
素材:@Image 1 是挑战者完整人物参考,保持原服装;不要继承图片背景。
故事:登场 → 滚筒失衡但通过 → 摆锤窄桥 → 成功冲线。
声音:普通话解说、现场欢呼、机关声和紧张BGM,无字幕。
输出:完整导演方案和最终提示词。
只要提示词
使用 $seedance-2-5-video-director。只要最终提示词,不要解释。
生成15秒、9:16的街头追逐戏:一名成年女性被三人追赶,写实手持动作片风格,人物与空间方向稳定。
诊断现有提示词
使用 $seedance-2-5-video-director。只诊断,不重写。
检查下面提示词的时间、运镜、人物、物理和声音是否互相冲突:<粘贴提示词>
更多模式与结构见 SKILL.md、提示词蓝图 和 首次使用引导。
人物参考原则
普通的“人物参考图”默认代表完整可见人物,而不是只参考脸:
- 保持身份、五官、脸型、肤色和年龄感
- 保持发型、发色、身高、体型和身体比例
- 保持可见服装、鞋子、配饰和穿着方式
- 默认不继承参考图背景、姿势、构图、裁切、文字和光线
- 只有用户明确说“只参考脸”或允许换装时,才缩小参考范围
目录
seedance-2-5-video-director/
├── SKILL.md
├── README.md
├── LICENSE
├── NOTICE.md
├── agents/
│ └── openai.yaml
├── references/
│ ├── capabilities-and-limits.md
│ ├── first-use-onboarding.md
│ ├── multimodal-patterns.md
│ ├── official-examples.md
│ ├── prompt-blueprints.md
│ └── realistic-direction-patterns.md
├── scripts/
│ └── validate_skill.py
├── tests/
│ └── cases.json
└── .github/workflows/
└── validate.yml
测试
本地运行:
python scripts/validate_skill.py
验证器会检查:
- Skill前置元数据和名称
- 必需文件及Markdown相对链接
- 安装引导触发规则与学习内容
agents/openai.yaml调用名称- 测试案例格式、唯一ID和主要模式覆盖
- 残留TODO或占位文本
GitHub Actions会在每次push和pull request时运行同一验证器。
tests/cases.json 是行为契约测试集。它不会调用视频生成服务,而是记录输入场景和预期输出约束,适合人工回归或交给独立代理做前向测试。
能力边界
- 本Skill只产出策划、脚本、诊断和提示词文本。
- 不自动提交视频生成任务,不消耗积分,不承诺生成结果。
- 平台上传限制以当前Dreamina/即梦界面为准;仓库中的能力摘要有快照日期。
- 生成模型存在随机性,提示词只能降低身份漂移、物理错误和镜头冲突,不能保证每次结果完全一致。
开源与第三方说明
代码与原创Skill内容采用 MIT License。平台名称、商标和第三方文档归各自权利人所有。能力摘要和示例来源说明见 NOTICE.md。
// HOW IT'S BUILT
// FILES
15 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.
SKILL.md14 KBb71f1b5d16cdupstream ↗
--- name: seedance-2-5-video-director description: Design, derive, optimize, diagnose, and rewrite scripts or copy-ready prompts specifically for Dreamina/即梦 Seedance 2.5. Use for new text/image/video/audio-to-video concepts, exact-timeline videos, 30-180-second Long Video, extension, video editing, Clay Renderer, seamless transitions, multi-grid storyboards, realistic acted dialogue, relationship drama, sports or variety-show live camera grammar, character and wardrobe locking, example-structure adaptation, voice/timbre reference, BGM removal, creative transfer, perspective reconstruction, and green-screen compositing. Produce only the requested deliverable—director brief, video script, diagnosis, revision, or final Seedance 2.5 prompt—and never submit a paid video-generation job. --- # Seedance 2.5 Video Director Translate a video idea, existing prompt, storyboard, or labeled image/video/audio set into a controllable Seedance 2.5 prompt. Direct the model-specific execution; do not replace genre-specific story development when another director Skill already supplies it. ## Installation onboarding protocol 1. Show [references/first-use-onboarding.md](references/first-use-onboarding.md) only when the installer or runtime explicitly indicates that the Skill was just installed, or when the user explicitly asks for help, installation guidance, a tutorial, the input template, or examples. 2. A new conversation or the Skill's first invocation in a conversation is not evidence of a new installation. Never repeat onboarding merely because conversation-level or cross-session state is unavailable. 3. For an ordinary video request, immediately perform the requested task without prepending onboarding content. 4. When installation onboarding is genuinely triggered, display its `学习指导`, `最小输入模板`, and `学习示例`, then continue the same request without asking the user to repeat it. 5. The onboarding guide teaches usage only. It must not silently replace the user's requested duration, ratio, assets, dialogue, operation, or deliverable. ## Required loading order 1. Apply the installation onboarding protocol above only when its explicit trigger is present. 2. Read [references/capabilities-and-limits.md](references/capabilities-and-limits.md) for every request. Validate media and duration before drafting. 3. Read [references/prompt-blueprints.md](references/prompt-blueprints.md) and load only the selected primary-mode section. 4. Read [references/multimodal-patterns.md](references/multimodal-patterns.md) whenever the request includes reference media or any specialist capability. 5. Read [references/realistic-direction-patterns.md](references/realistic-direction-patterns.md) whenever the request includes realistic human acting, dialogue, emotional conflict, a relationship scene, sports/variety live coverage, obstacle physics, or a user-supplied example whose structure should be adapted. 6. Read [references/official-examples.md](references/official-examples.md) only when a concrete provider example is needed to resolve an underspecified structure, optimize an existing prompt, or check a difficult mode. Do not copy an example mechanically. ## Boundary - Produce planning and text prompts only. Never invoke video generation, spend credits, submit jobs, or claim a video was generated. - Keep this Skill provider-specific. Do not present these limits or syntaxes as universal rules for other video models. - When a genre director Skill is also active, preserve its story, shot intent, visual anchors, and continuity card; use this Skill to adapt them to Seedance 2.5 modes, asset references, timing, and prompt syntax. ## Select one primary mode Use explicit user wording first, then infer from duration, source media, and requested operation. Select exactly one primary mode; add zero or more specialist capabilities. | Primary mode | Select when | |---|---| | `basic-multimodal` | Create a new 4-30 second video without a stricter mode below | | `timestamp-30s` | Create a new video with an exact 30-second timeline or fine timestamp control | | `long-video` | Generate a new 30-180 second video in Long Video mode | | `video-extension` | Extend an existing video forward or backward | | `video-edit` | Modify an existing generated/uploaded video with Smart Edit, Edit with marks, or Edit Video | | `clay-renderer` | Use a coarse or fine 3D white-model/Clay Renderer video as motion, composition, or rendering control | | `seamless-transition` | Preserve two source videos and generate only a connecting bridge | | `multi-grid-storyboard` | Animate a multi-panel storyboard where each panel represents a planned shot | Treat realistic-person direction, multiple-person locking, timbre reference, BGM removal, creative transfer, local removal/replacement, perspective reconstruction, and green-screen compositing as specialist capabilities, not extra primary modes. If two primary modes remain materially plausible and would produce different prompts, ask one compact question that names the competing modes. Do not ask for information that can be safely inferred. ## Minimal-question policy - Preserve every explicit user parameter. Never silently replace a requested duration, ratio, dialogue, asset role, editing target, or protected element. - Lock the requested deliverable before drafting. `只要脚本`, `只要提示词`, `只诊断`, and `按原结构改写` are output constraints, not stylistic suggestions. - Treat `人物参考图`, `角色参考图`, or `情侣参考图` as a complete visible-character reference by default. Apply the detailed scope rules in `multimodal-patterns.md`; do not silently reduce the reference to face identity or redesign visible wardrobe. - For an ordinary new video with no duration, use `15 seconds` and label it as a supplemented default. - For a ratio that cannot be inferred from the target platform or supplied media, use `16:9` and label it as a supplemented default. - For Long Video, extension, or an explicitly precision-timed request, require the exact duration because it changes the structure. - Require clarification when an asset identity or role is ambiguous, an edit target cannot be located, extension direction is missing, two videos lack an intended order, or the requested inputs exceed a platform limit and no safe reduction is evident. - Otherwise, make restrained defaults and proceed. Group all indispensable questions into one turn. ## Direction workflow 1. **Lock the request.** Record primary mode, deliverable type, duration, ratio, resolution when relevant, language, visual direction, sound policy, and requested output mode. Mark inferred defaults. 2. **Inventory assets.** Use only media the user actually supplied or explicitly numbered. Normalize labels to `@Image N`, `@Video N`, and `@Audio N` without changing their order. 3. **Build the asset lock table.** For every asset, record `label | role | active time | preserve | do not inherit`. Omit the table only when there are no assets. 4. **Lock example inheritance.** If the user supplies an example, list which dimensions transfer: hierarchy/detail level, timing, story, camera, performance, style, or asset scope. `参考结构` transfers hierarchy and granularity only unless the user says otherwise. 5. **Assign specialist capabilities.** Add only capabilities the request needs. Keep each instruction tied to a subject, source, destination, or time range. 6. **Choose the blueprint and control depth.** Follow the selected mode in [references/prompt-blueprints.md](references/prompt-blueprints.md). Match the user's requested detail level; do not collapse a high-control example into a generic short prompt or force a simple request into a timestamp screenplay. 7. **Write continuity locks.** Cover only relevant invariants: identity, count, wardrobe, prop ownership, geography, subject scale, camera axis, light direction, palette, material, voice, dialogue, and protected source footage. 8. **Draft the requested artifact.** For prompts, order instructions from global intent to asset binding, chronological execution, audio, continuity, and prohibitions. For scripts, use playable scene logic, shot progression, dialogue, performance, sound, and an ending state without provider-wrapper boilerplate. 9. **Audit.** Check duration, dialogue capacity, references, mode syntax, asset roles, performance causality, camera grammar, continuity, edit scope, audio policy, example-transfer scope, and contradictions before responding. ## Timing and reference rules - Make timestamp segments continuous, non-overlapping, and collectively equal to the requested duration. Use one consistent boundary convention such as `0-5s, 5-10s`. - Give each segment enough time for its actions. Reduce event count before compressing actions into implausibly short intervals. - Audit spoken lines against available performance time. Leave room for listening, comprehension, breath, interruption, and reaction; shorten dialogue before forcing unnatural delivery. - Repeat a reference only where it matters. State what to borrow and what not to borrow; `reference @Video 1` alone is insufficient. - Bind each person to one identity source and keep clothing, face, body proportions, and role independent from other people. - Keep camera movement physically compatible with framing changes and subject motion. Prefer one primary move per beat. - Make emotional and physical results causal. A face reacts after the line is heard; a body moves after contact or force; tears, laughter, blush, falling, and recovery must not appear before their trigger. - Put dialogue, sound effects, ambience, timbre sources, BGM policy, and subtitle policy in explicit audio instructions. - Do not invent asset labels, UI controls, source content, dialogue, or visual details the user has not supplied unless they are clearly labeled creative supplements. ## Mode-specific invariants - **Extension:** normalize direction to `prepend-before-source` or `append-after-source`; never rely on ambiguous phrases such as `向前续写`, `前向续写`, `往前延长`, `向后续写`, or `后向续写` without an explicit before/after boundary. For prepend, make the added interval end at and hand off into the source's first frame. For append, make it begin from the source's last frame. Describe only the newly generated interval; state placement and added duration; preserve the original interval; require natural action, camera, lighting, sound, and spatial connection; prohibit unexplained resets and objects appearing from nowhere. Unless the user explicitly says the added interval is before the original, after the original, connected to the first frame, or connected to the last frame, ask `新增片段放在原片之前,还是原片之后?` and do not draft the prompt. Do not infer placement from the requested plot alone. - **Video edit:** use `annotation/location + exact target + add/remove/replace/change + effective time`; explicitly list everything that must remain unchanged. For an unmarked edit, omit the annotation clause. - **Clay Renderer:** classify coarse versus fine. For coarse references, borrow motion, camera, position, rhythm, or lighting changes and map every model to its final subject. For fine references, issue a render command and describe materials, environment, light, and any timed rendering changes. Remove trajectory lines, axes, and camera cones from final imagery. - **Seamless transition:** state that both source videos must remain unmodified; define the bridge trigger and lock the connection object's shape, position, scale, direction, speed, motion trend, camera continuity, and visual transformation. - **Multi-grid storyboard:** declare that each panel is a complete ordered shot; map character and scene references; describe every shot's time, composition, action, and camera; preserve spatial and identity continuity between panels. ## Output modes Honor the most specific requested deliverable: - `full-direction` (default): return `导演方案`, `素材映射`, `连续性与禁止项`, and exactly one copy-ready `Seedance 2.5 最终提示词` block. - `prompt-only`: return only one fenced `text` block; omit headings, diagnosis, mapping, and commentary. - `script-only`: return only the playable video script. Include the premise, scene/character psychology when useful, continuous timing or a shot list, exact dialogue, visible performance, camera, sound, and ending. Omit the asset table and Seedance prompt wrapper unless requested. - `diagnosis-only`: explain the cause, cite specific contradictions or weak controls, and do not silently rewrite the whole artifact. - `revision`: preserve the requested source structure and unchanged elements; summarize material changes briefly, then provide the revised artifact in the user's requested format. Write in Chinese by default. Preserve exact dialogue language, asset labels, proper nouns, and requested output language. ## Final audit Before responding, verify: - one primary mode is selected and compatible with duration; - all referenced assets exist and their labels match the supplied order; - complete-character references preserve visible wardrobe and body traits unless the user explicitly narrows the scope; - no asset has two conflicting identity or purpose assignments; - example-based work inherits only the explicitly requested dimensions; - timestamps cover the total exactly with no gap or overlap; - spoken dialogue, pauses, actions, and reactions plausibly fit their intervals; - extension text affects only the new interval; - edit target, operation, location, and active time are explicit; - seamless-transition text protects both source videos and defines a bridge; - character identities, counts, props, and voices do not cross-bind; - POV, third-person view, one-take/cut grammar, shot scale, same-frame rules, screen direction, and eyelines agree; - realistic performance uses motivated transitions rather than mechanical facial checklists or unearned blush/tears; - physical action shows cause, contact or force, body/object response, and outcome in the correct order; - sound, BGM, subtitle, and dialogue requirements agree; - the final answer contains no promise or action of direct video generation.README.md6.0 KB9427d6bb009aupstream ↗
# Seedance 2.5 Video Director 面向 Dreamina/即梦 Seedance 2.5 的视频导演 Skill。它将文本创意、人物参考图、视频、音频、分镜或旧提示词整理成可执行的视频脚本、导演方案、诊断结果和复制即用的提示词。 > 非字节跳动、Dreamina 或即梦官方项目。本仓库不会调用付费生成接口,不会自动消耗积分。 ## 功能 - 4–30秒普通视频与精确30秒时间轴 - 30–180秒 Long Video - 视频前置/后置续写 - Smart Edit、标记区域编辑和上传视频编辑 - Clay Renderer 粗白模/精白模渲染指令 - 两段视频无缝转场 - 多格分镜动画 - 人物身份、体型、服装和配饰锁定 - 多人物、对白音色、BGM移除与声画同步 - 真人情感表演、情侣戏、竞技综艺直播镜头和障碍物物理 - 提示词冲突诊断与按原结构改写 ## 安装 ### 推荐:一行安装 已安装 Node.js 的用户,在 Windows、macOS 或 Linux 终端执行: ```bash npx skills add liyue-aigc/seedance-2-5-video-director -g -y ``` 这条命令会从 GitHub 获取 Skill,并安装到当前用户的 Agent Skills 目录;`-g` 表示全局安装,`-y` 表示跳过交互确认。它使用 npm 附带的 `npx` 启动安装器,但本仓库本身不是需要加入项目依赖的 npm 软件包,因此不使用 `npm install`。 如果电脑上没有 Node.js,也可以直接在 Codex 中发送: ```text 请使用 $skill-installer,从 GitHub 安装 liyue-aigc/seedance-2-5-video-director。 ``` ### 手动安装(备用) #### Windows PowerShell ```powershell $skillRoot = Join-Path $env:USERPROFILE '.codex\skills' New-Item -ItemType Directory -Force -Path $skillRoot | Out-Null git clone https://github.com/liyue-aigc/seedance-2-5-video-director.git (Join-Path $skillRoot 'seedance-2-5-video-director') ``` #### macOS / Linux ```bash mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills" git clone https://github.com/liyue-aigc/seedance-2-5-video-director.git \ "${CODEX_HOME:-$HOME/.codex}/skills/seedance-2-5-video-director" ``` 安装完成后重新打开Codex会话,并在提示词中调用: ```text 使用 $seedance-2-5-video-director,把我的视频创意整理成可复制的 Seedance 2.5 提示词。 ``` 安装器明确提示刚完成安装时,Skill会展示一次学习指导;普通新对话会直接处理任务,不会重复教程。需要时可主动发送“显示 `$seedance-2-5-video-director` 使用教程”。 ## 更新 通过 `npx skills` 安装的用户: ```bash npx skills update seedance-2-5-video-director -g -y ``` 手动克隆的用户: ```bash git -C "${CODEX_HOME:-$HOME/.codex}/skills/seedance-2-5-video-director" pull ``` Windows默认安装路径更新: ```powershell git -C (Join-Path $env:USERPROFILE '.codex\skills\seedance-2-5-video-director') pull ``` ## 快速使用 ```text 使用 $seedance-2-5-video-director。 任务:生成20秒、16:9的水上竞技综艺片段。 素材:@Image 1 是挑战者完整人物参考,保持原服装;不要继承图片背景。 故事:登场 → 滚筒失衡但通过 → 摆锤窄桥 → 成功冲线。 声音:普通话解说、现场欢呼、机关声和紧张BGM,无字幕。 输出:完整导演方案和最终提示词。 ``` ### 只要提示词 ```text 使用 $seedance-2-5-video-director。只要最终提示词,不要解释。 生成15秒、9:16的街头追逐戏:一名成年女性被三人追赶,写实手持动作片风格,人物与空间方向稳定。 ``` ### 诊断现有提示词 ```text 使用 $seedance-2-5-video-director。只诊断,不重写。 检查下面提示词的时间、运镜、人物、物理和声音是否互相冲突:<粘贴提示词> ``` 更多模式与结构见 [SKILL.md](SKILL.md)、[提示词蓝图](references/prompt-blueprints.md) 和 [首次使用引导](references/first-use-onboarding.md)。 ## 人物参考原则 普通的“人物参考图”默认代表完整可见人物,而不是只参考脸: - 保持身份、五官、脸型、肤色和年龄感 - 保持发型、发色、身高、体型和身体比例 - 保持可见服装、鞋子、配饰和穿着方式 - 默认不继承参考图背景、姿势、构图、裁切、文字和光线 - 只有用户明确说“只参考脸”或允许换装时,才缩小参考范围 ## 目录 ```text seedance-2-5-video-director/ ├── SKILL.md ├── README.md ├── LICENSE ├── NOTICE.md ├── agents/ │ └── openai.yaml ├── references/ │ ├── capabilities-and-limits.md │ ├── first-use-onboarding.md │ ├── multimodal-patterns.md │ ├── official-examples.md │ ├── prompt-blueprints.md │ └── realistic-direction-patterns.md ├── scripts/ │ └── validate_skill.py ├── tests/ │ └── cases.json └── .github/workflows/ └── validate.yml ``` ## 测试 本地运行: ```bash python scripts/validate_skill.py ``` 验证器会检查: - Skill前置元数据和名称 - 必需文件及Markdown相对链接 - 安装引导触发规则与学习内容 - `agents/openai.yaml` 调用名称 - 测试案例格式、唯一ID和主要模式覆盖 - 残留TODO或占位文本 GitHub Actions会在每次push和pull request时运行同一验证器。 `tests/cases.json` 是行为契约测试集。它不会调用视频生成服务,而是记录输入场景和预期输出约束,适合人工回归或交给独立代理做前向测试。 ## 能力边界 - 本Skill只产出策划、脚本、诊断和提示词文本。 - 不自动提交视频生成任务,不消耗积分,不承诺生成结果。 - 平台上传限制以当前Dreamina/即梦界面为准;仓库中的能力摘要有快照日期。 - 生成模型存在随机性,提示词只能降低身份漂移、物理错误和镜头冲突,不能保证每次结果完全一致。 ## 开源与第三方说明 代码与原创Skill内容采用 [MIT License](LICENSE)。平台名称、商标和第三方文档归各自权利人所有。能力摘要和示例来源说明见 [NOTICE.md](NOTICE.md)。scripts/validate_skill.py6.8 KB00611cfba4fbupstream ↗
#!/usr/bin/env python3 """Deterministic structural validation for the public Skill repository.""" from __future__ import annotations import json import re import sys from pathlib import Path ROOT = Path(__file__).resolve().parents[1] SKILL_NAME = "seedance-2-5-video-director" def fail(message: str, errors: list[str]) -> None: errors.append(message) def validate_required_files(errors: list[str]) -> None: required = [ "SKILL.md", "README.md", "LICENSE", "NOTICE.md", "agents/openai.yaml", "references/capabilities-and-limits.md", "references/first-use-onboarding.md", "references/multimodal-patterns.md", "references/official-examples.md", "references/prompt-blueprints.md", "references/realistic-direction-patterns.md", "tests/cases.json", ".github/workflows/validate.yml", ] for relative in required: if not (ROOT / relative).is_file(): fail(f"missing required file: {relative}", errors) def validate_skill_frontmatter(errors: list[str]) -> None: text = (ROOT / "SKILL.md").read_text(encoding="utf-8") if not text.startswith("---\n"): fail("SKILL.md must start with YAML frontmatter", errors) return parts = text.split("---", 2) if len(parts) < 3: fail("SKILL.md frontmatter is not closed", errors) return frontmatter = parts[1] name_match = re.search(r"(?m)^name:\s*(.+?)\s*$", frontmatter) description_match = re.search(r"(?m)^description:\s*(.+?)\s*$", frontmatter) if not name_match or name_match.group(1).strip('"\'') != SKILL_NAME: fail(f"SKILL.md name must be {SKILL_NAME}", errors) if not description_match or len(description_match.group(1).strip()) < 80: fail("SKILL.md description is missing or too short", errors) required_phrases = [ "Installation onboarding protocol", "references/first-use-onboarding.md", "A new conversation or the Skill's first invocation in a conversation is not evidence of a new installation", ] for phrase in required_phrases: if phrase not in text: fail(f"SKILL.md missing first-use contract: {phrase}", errors) def validate_onboarding(errors: list[str]) -> None: text = (ROOT / "references/first-use-onboarding.md").read_text(encoding="utf-8") for heading in ("## 学习指导", "## 最小输入模板", "## 学习示例"): if heading not in text: fail(f"onboarding missing section: {heading}", errors) if text.count("### 示例") < 4: fail("onboarding must contain at least four learning examples", errors) if f"${SKILL_NAME}" not in text: fail("onboarding minimal template must invoke the Skill by name", errors) if "当前会话首次调用时" in text: fail("onboarding must not trigger merely on the first invocation of a conversation", errors) def validate_markdown_links(errors: list[str]) -> None: link_pattern = re.compile(r"\[[^\]]+\]\(([^)]+)\)") for markdown in ROOT.rglob("*.md"): text = markdown.read_text(encoding="utf-8") for target in link_pattern.findall(text): target = target.strip().split("#", 1)[0] if not target or "://" in target or target.startswith("#"): continue resolved = (markdown.parent / target).resolve() try: resolved.relative_to(ROOT.resolve()) except ValueError: fail(f"link escapes repository: {markdown.relative_to(ROOT)} -> {target}", errors) continue if not resolved.exists(): fail(f"broken link: {markdown.relative_to(ROOT)} -> {target}", errors) def validate_agent_metadata(errors: list[str]) -> None: text = (ROOT / "agents/openai.yaml").read_text(encoding="utf-8") if f"${SKILL_NAME}" not in text: fail("agents/openai.yaml default_prompt must mention the Skill", errors) if "display_name:" not in text or "short_description:" not in text: fail("agents/openai.yaml is missing interface metadata", errors) def validate_cases(errors: list[str]) -> None: path = ROOT / "tests/cases.json" try: cases = json.loads(path.read_text(encoding="utf-8")) except (OSError, json.JSONDecodeError) as exc: fail(f"invalid tests/cases.json: {exc}", errors) return if not isinstance(cases, list) or len(cases) < 10: fail("tests/cases.json must contain at least ten cases", errors) return ids: set[str] = set() categories: set[str] = set() for index, case in enumerate(cases): if not isinstance(case, dict): fail(f"case {index} must be an object", errors) continue for field in ("id", "category", "request", "expected"): if field not in case: fail(f"case {index} missing field: {field}", errors) case_id = case.get("id") if isinstance(case_id, str): if case_id in ids: fail(f"duplicate case id: {case_id}", errors) ids.add(case_id) category = case.get("category") if isinstance(category, str): categories.add(category) expected = case.get("expected") if not isinstance(expected, list) or not expected: fail(f"case {case_id or index} expected must be a non-empty list", errors) required_categories = { "onboarding", "basic-multimodal", "timestamp-30s", "long-video", "video-extension", "video-edit", "clay-renderer", "seamless-transition", "multi-grid-storyboard", "diagnosis", } missing = sorted(required_categories - categories) if missing: fail(f"test categories missing: {', '.join(missing)}", errors) def validate_no_placeholders(errors: list[str]) -> None: placeholder = re.compile(r"\b(TODO|TBD|FIXME)\b", re.IGNORECASE) for path in [ROOT / "SKILL.md", ROOT / "README.md", *ROOT.joinpath("references").glob("*.md")]: match = placeholder.search(path.read_text(encoding="utf-8")) if match: fail(f"placeholder {match.group(0)} found in {path.relative_to(ROOT)}", errors) def main() -> int: errors: list[str] = [] validate_required_files(errors) if not errors: validate_skill_frontmatter(errors) validate_onboarding(errors) validate_markdown_links(errors) validate_agent_metadata(errors) validate_cases(errors) validate_no_placeholders(errors) if errors: print("Validation failed:") for error in errors: print(f"- {error}") return 1 print("Validation passed: repository structure, onboarding, links, metadata, and test contracts are valid.") return 0 if __name__ == "__main__": sys.exit(main())references/capabilities-and-limits.md4.6 KB6cdfca3904d9upstream ↗
# Seedance 2.5 capabilities and limits ## Contents - Source snapshot - Mode limits - Input requirements - Stability guidance - Validation behavior ## Source snapshot - Source: `https://bytedance.larkoffice.com/wiki/NjnWwvf4BiFYFLk2RzrcEgaunGf` - Document title: `【Dreamina】Seedance 2.5 User Guide` - Document revision: `419` - Snapshot organized: `2026-08-16` Treat this local file as the V1 runtime authority so the Skill does not depend on network access or Lark permissions. If the Dreamina UI conflicts with this snapshot, state the discrepancy and follow the current UI for submission constraints; do not silently invent updated limits. ## Mode limits | Mode | Duration or source rule | Notes | |---|---|---| | Basic, first frame, first/last frame, Omni Reference | 4-30 seconds output | Handbook also expresses this as 97-721 frames; output resolution listed as 480p or 720p | | Long Video | 30-180 seconds output | Select Long Video instead of segment-by-segment extension | | Video extension | Source video must be under 30 seconds | Describe only the new interval; handbook advertises nested continuation up to 60 seconds total | | Smart Edit / Edit with marks / Edit Video | Prefer source videos within 20 seconds for stability | Longer sources may work with reduced stability | | Seamless transition | Two ordered source videos | Keep both originals unchanged and generate the gap/bridge | | Clay Renderer | Coarse or fine white-model reference video | Coarse is currently described as the more stable generation path | Do not turn an ordinary `30 seconds` request into Long Video automatically. Use `timestamp-30s` for precise 30-second generation unless the user explicitly asks for Long Video behavior. ## Input requirements ### Images - Formats: JPEG, PNG, WebP, BMP, TIFF, GIF, HEIC, HEIF. - Aspect ratio `width / height`: greater than 0.4 and less than 2.5. - Width and height: 300-6000 px. - Single-image size: less than 30 MB. - Request body: at most 64 MB; avoid Base64 for large files. - Seedance 2.5 maximum: 30 images per request; handbook describes up to 4K for the maximum-count workflow. ### Videos - Formats: MP4, MOV. - Resolution: 480p-4K. - Aspect ratio `width / height`: 0.4-2.5. - Width and height: 300-6000 px. - Total pixels: 409,600-8,295,044. - Single-video size: at most 200 MB. - Frame rate: 24-60 FPS. - Single-video nominal duration: 2-30 seconds; implementation tolerance is described as 1.8-30.2 seconds. - Maximum: 10 reference videos with a combined nominal duration of at most 30 seconds; implementation tolerance is described as 30.2 seconds. ### Audio - Formats: WAV, MP3. - Single-audio size: at most 15 MB. - Request body: at most 64 MB; avoid Base64 for large files. - Single-audio nominal duration: 2-30 seconds; implementation tolerance is described as 1.8-30.2 seconds. - Maximum: 10 reference audio clips with a combined nominal duration of at most 30 seconds; implementation tolerance is described as 30.2 seconds. - Audio-only reference input is supported in Seedance 2.5. ## Stability guidance - Main subjects driven by reference video or audio: 1-5 is preferred; 6-10 may reduce stability. - Main reference video/audio duration: 5-10 seconds is preferred; longer inputs may reduce stability. - Subjects in still images: 1-8 is preferred; 9-12 may reduce stability. - When all character identities are supplied through still images, apply the still-image guidance above: eight subjects remain within the preferred range and should not trigger a stability warning merely because the cast exceeds five. - For more than five subjects, separate multi-view references into multiple images rather than one image containing many views. Stability order: multiple images with one view each is better than one image containing multiple views. - Video edit source: within 20 seconds is preferred. - Reference images for video editing: 1-5 is preferred; 6-8 may reduce stability. Treat these as quality recommendations, not hard rejections. ## Validation behavior 1. Count actual or explicitly declared assets. Never count hypothetical placeholders. 2. Check format, individual duration, combined duration, size, ratio, and count when metadata is available. 3. If a hard limit is exceeded, stop prompt drafting and propose a concrete reduction, split, trim, re-encode, or alternate mode. 4. If metadata is unavailable, do not fabricate it. Add a short preflight reminder only when the missing metadata could block submission. 5. If the request is within hard limits but beyond a stability recommendation, continue and label the risk with a practical simplification option. 6. Never claim that a prompt can override an upload or UI limit.references/first-use-onboarding.md4.6 KB6db3aff9aa18upstream ↗
# Seedance 2.5 视频导演安装与使用引导 ## 输出规则 仅在安装器或运行环境明确说明 Skill 刚完成安装,或用户主动索要帮助、安装指导、教程、输入模板或示例时展示。不要因为新开对话、当前对话第一次调用或缺少跨会话状态而自动展示。若展示后仍有实际任务,立即继续处理原请求,不要求用户重新发送。 ## 学习指导 ```text 【Seedance 2.5 视频导演|使用指导】 这个 Skill 的工作是把视频创意、人物参考图、已有视频、音频、分镜图或旧提示词,整理成适合 Dreamina/即梦 Seedance 2.5 的视频脚本、导演方案、诊断结果或可复制提示词。 它只负责策划和编写文本,不会自动提交生成任务、消耗积分或声称已经生成视频。 开始前只要确认五件事: 1. 想做什么:新视频、续写、编辑、转场、白模渲染或分镜动画; 2. 总时长和画幅:例如15秒、16:9; 3. 素材角色:每张图片、视频和音频分别控制谁、什么动作或什么声音; 4. 故事变化:从第一帧到最后一帧发生什么; 5. 输出形式:完整导演方案、只要提示词、只要脚本、只诊断或按原结构修改。 常用主模式: - 4–30秒普通新视频:basic-multimodal; - 精确30秒时间轴:timestamp-30s; - 30–180秒长视频:long-video; - 在原视频之前或之后增加内容:video-extension; - 删除、替换或修改原视频局部:video-edit; - 白模动作或场景渲染:clay-renderer; - 连接两段原视频:seamless-transition; - 把多格分镜依次做成镜头:multi-grid-storyboard。 人物参考图默认锁定完整可见人物,包括身份、脸、发型、体型、服装、鞋子和配饰。除非你明确说“只参考脸”或允许重新造型,否则不会自行换衣服。 续写必须说清新增片段位于原片之前还是之后;“向前续写”和“向后续写”容易产生歧义,Skill会先确认边界。 写真人表演时,优先描述人物为何反应、听见后的停顿、自然动作和残留情绪,不机械罗列眉毛、眼睛、鼻子和嘴的逐项变化。 ``` ## 最小输入模板 ```text 使用 $seedance-2-5-video-director。 任务:<新视频 / 续写 / 编辑 / 转场 / 白模渲染 / 分镜动画> 时长与画幅:<例如15秒,16:9> 素材:<例如 @Image 1 是完整人物参考;@Video 1 只参考动作> 故事:<开始状态 → 主要变化 → 最终状态> 声音:<对白、环境声、BGM、是否需要字幕> 输出:<完整方案 / 只要提示词 / 只要脚本 / 只诊断 / 按原结构修改> ``` ## 学习示例 ### 示例一:人物参考图生成新视频 ```text 使用 $seedance-2-5-video-director。 任务:生成20秒、16:9的第三人称情感短片。 素材:@Image 1 是女主完整人物参考,严格保持脸、发型、体型和原服装;不要继承图片背景。 故事:女主在雨停后的公交站等人,对方始终没来;她看完手机消息,收起伞独自离开。 声音:只有城市环境声和一句自然对白,无字幕,无BGM。 输出:完整导演方案和最终提示词。 ``` 学习重点:完整人物参考不等于只参考脸;时间轴必须给等待、理解消息和离开留出真实表演时间。 ### 示例二:原视频之后续写 ```text 使用 $seedance-2-5-video-director。 任务:在 @Video 1 原片之后追加5秒。 新增剧情:承接原片最后一帧的人物步伐、朝向、机位和环境声;人物停下,回头看向身后的门,门内灯光熄灭。 保护:原视频内容完全不变,不重置站位,不新增无关人物。 输出:只要最终提示词。 ``` 学习重点:明确“原片之后”和“最后一帧”,避免续写方向歧义。 ### 示例三:诊断矛盾提示词 ```text 使用 $seedance-2-5-video-director。 只诊断下面提示词,不要重写: “一镜到底,固定机位,同时切换正反打;人物始终距离镜头两米,但最后脸部填满画面。” ``` 学习重点:一镜到底与正反打冲突;固定距离与脸部填满画面冲突。诊断模式只说明问题,不擅自改写成另一个故事。 ### 示例四:视频编辑 ```text 使用 $seedance-2-5-video-director。 任务:编辑 @Video 1。 在5–8秒,仅把红色方框内的蓝色杯子替换为白色花瓶;人物、动作、桌面其他物品、背景、机位、原声和其余时间保持不变。 输出:只要最终提示词。 ``` 学习重点:编辑任务要同时写清时间、位置、目标、操作和不允许变化的元素。references/multimodal-patterns.md9.1 KB655589e30350upstream ↗
# Multimodal reference and specialist patterns ## Contents - Asset lock table - Reference grammar - Person-reference scope - Realistic people - Multiple people - Timbre and dialogue - BGM removal - Creative transfer - Local removal and replacement - Perspective reconstruction - Green-screen compositing - Conflict handling ## Asset lock table Before drafting, normalize every supplied or explicitly numbered asset into this internal/public table: | Asset | Role | Active time | Preserve | Do not inherit | |---|---|---|---|---| | `@Image 1` | identity / scene / prop / style / storyboard | global or time range | protected traits | irrelevant background, text, layout, etc. | | `@Video 1` | source / motion / camera / edit / transition / white model | global or time range | selected motion or footage | unwanted identity, material, audio, overlays, etc. | | `@Audio 1` | timbre / dialogue delivery / BGM / rhythm / ambience | global or time range | selected acoustic traits | noise, unrelated speech, unwanted music, etc. | Use only rows backed by actual user inputs. If the user describes roles but does not label upload order, preserve the app's attachment order and state the mapping. ## Reference grammar Write each reference as a scoped contract: `Reference source + traits to borrow + target subject/shot + active time + protected invariants + traits not to inherit` Good: `20-25 秒仅参考 @Video 2 中“预备—冲刺—起跳—空中举剑—挥砍”的重心和动作节奏,将动作映射给 @Image 1 的小骑士;不要继承 @Video 2 的灰模人物、场景材质或其他时间段动作。` Weak: `参考 @Video 2 的动作。` Separate roles when one asset serves multiple purposes. Do not assume that a motion reference should also control character identity, background, audio, or style. ## Person-reference scope Interpret ordinary person-reference language conservatively: - `人物参考图`, `角色参考图`, `人物视觉参考`, `情侣参考图`, or `严格参考人物图片` defaults to the complete visible character: identity, facial features, face shape, skin tone, age impression, hair shape/color, stature, body frame and proportions, visible wardrobe, footwear, accessories, and overall demeanor. - Do not redesign, replace, recolor, add, or remove visible wardrobe or accessories unless the user explicitly permits styling changes. - Reconstruct unseen sides or backs consistently from visible cut, color, material, construction, and wearing method; do not invent a different outfit. - By default, do not inherit the source image's background, composition, pose, text, crop, or lighting unless the user assigns those roles too. - If the user says `只参考脸`, `只锁身份`, or otherwise narrows the role, preserve only that declared scope and label any wardrobe or styling as a creative supplement. - If one image contains a couple or group, bind each visible person independently by position or user-named role. Preserve relative height and scale; prohibit face mixing, wardrobe swapping, duplicated bodies, and role reversal. If the image itself is unavailable for inspection but the user declares it as a person reference, keep the reference contract symbolic and preserve the visible character. Do not compensate for missing inspection by inventing clothing. ## Realistic people Use this structure only to the level needed by the shot: `specific adult age and ethnicity + skin tone and real texture + 3-4 distinctive facial traits + gaze/emotional information + hairstyle/color/state + garment cut/color/material/wear + body frame/posture/temperament` Reality anchors: - Retain micro-pores, fine skin texture, natural asymmetry, and context-appropriate minor marks; avoid flat plastic skin. - Describe eyes through visible gaze behavior and underlying emotion, not empty adjectives alone. - Give hair a cut, texture, condition, and environmental response. - Give clothing a cut, material, condition, and wearing detail. - Lock age, face, hair silhouette, garment, body proportions, and identity across all shots. - Default text-only people to fictional, clearly adult subjects. Preserve a real person's identity only from authorized user-provided references. Do not overload distant or fast shots with invisible facial microdetail. For acted dialogue or emotional realism, load `realistic-direction-patterns.md`. Prefer motivated visible behavior over anatomy-by-anatomy expression instructions. ## Multiple people Bind every important person independently: `Character A = @Image 1; Character B = @Image 2; ...` For each character, lock: - identity and facial features; - hairstyle and clothing; - body proportions and scale; - spatial position or entrance direction; - owned props; - action, gaze target, and relationship; - voice/timbre source when dialogue exists. State counts explicitly for groups. Keep roles primary/secondary. During physical interaction, describe contact and occlusion so limbs do not merge. Add `no face mixing, wardrobe swapping, duplicated bodies, extra limbs, or prop transfer` only when relevant. ## Timbre and dialogue - Prefer clear, noise-free human-voice reference audio. - Bind each voice to one character and quote the exact line. - State which qualities to retain: pitch/color, emotional tension, pace, pauses, breath, intensity, or whisper/shout behavior. - Scope the timbre to the relevant line or time range. - Do not use a single vague instruction such as `use @Audio 1`. Write: `角色 A 在 8-11 秒说“……”,音色参考 @Audio 1,保持低沉、克制、略带停顿的悬疑语气。` - Keep dialogue duration plausible. Shorten the line or expand the beat when speech cannot fit naturally. ## BGM removal For lossless dialogue extraction, write: `移除 @Video N 的背景音乐,仅保留清晰的人声对白和指定现场声;画面、原有字幕及其他受保护视觉元素保持不变。` Name which non-vocal sounds to preserve or remove. Distinguish: - `remove BGM, retain dialogue and ambience`; - `mute all sound`; - `retain only dialogue`; - `replace BGM with @Audio N`. Do not combine `no BGM` with a contradictory music request. ## Creative transfer Transfer creative logic, not the entire source: - Identify camera rhythm, motion grammar, transition method, emotional arc, comedic/cute state, composition logic, color progression, or social-media pacing to borrow. - Map the transferred logic to the target identity, scene, props, and style. - Lock target appearance and core content. - Explicitly reject unwanted source identity, background, text, logo, audio, or texture. Pattern: `参考 @Video 1 的 [creative traits],将其转译到 @Image 1 的 [target] 与 [new setting];保持 [target invariants],不要继承 [source exclusions]。` ## Local removal and replacement Use this as a specialist capability within `video-edit`: - Location: marked region, screen side, named object, or relationship to another object. - Target: exact person/object/background element. - Operation: remove, add, or replace A with B. - Time: global, fixed interval, or gradual onset. - Reconstruction: fill the background using original geometry, parallax, occlusion, motion, material, and light. - Protection: list the subjects, actions, framing, background, sound, and text that remain unchanged. For marked edits, specify `inside/outside the red box`, `at the blue landmark`, `where the arrow points`, or the exact tool and direction. Never rely on `change this`. ## Perspective reconstruction State source and destination camera explicitly: `Change [source view/framing] to [target view/framing/movement], while preserving [subjects, action timing, space, background, and identity].` Control: - camera height and angle: overhead, eye level, low angle; - viewpoint: first person, third person, fixed frontal; - framing/lens: close-up, medium, wide, focal behavior; - movement: push, pull, pan, tilt, orbit, track; - reconstructed off-screen space and scale logic. If a reference video supplies the target perspective, state what perspective traits to borrow and what original content not to inherit. ## Green-screen compositing Map foreground, background, and effects separately: - identify the green-screen source and the subject to extract; - identify the destination scene or plate; - preserve foreground contour, motion, timing, and identity; - match camera angle, scale, ground contact, parallax, light direction, color spill, shadow, atmosphere, and motion blur; - define interaction timing for explosions, weather, particles, or passing objects; - remove green spill, edge halos, flicker, floating feet, and pasted-on lighting. Do not call a non-green-screen replacement `green-screen compositing` unless the user requests that workflow. ## Conflict handling Resolve instructions in this order: 1. Explicit protected source footage or identity. 2. Explicit edit target and time range. 3. Explicit asset role. 4. Explicit duration, dialogue, audio, and subtitle requirements. 5. Style and creative supplements. If two explicit instructions conflict at the same priority, ask one question rather than choosing silently. If a lower-priority style reference conflicts with an identity or protected-footage lock, preserve the lock and use only compatible style traits.references/official-examples.md9.6 KB63d43f807227upstream ↗
# Curated handbook examples ## Contents - Provenance and use - Basic multimodal - Exact 30-second structure - Long Video - Video extension - Video edit - Clay Renderer - Seamless transition - Multi-grid storyboard - Specialist snippets ## Provenance and use - Source: `https://bytedance.larkoffice.com/wiki/NjnWwvf4BiFYFLk2RzrcEgaunGf` - Source revision: `419` - Curated: `2026-08-16` These examples are normalized and shortened from the handbook to expose reusable structure. They are not an exhaustive mirror. Use them to understand specificity and ordering; replace all subjects, references, timing, and constraints with the user's request. ## Basic multimodal ```text 写实自然纪录片风格,电影级真实光影。清晨薄雾中的秋季湿地芦苇荡参考 @Image 1;红冠鹤的羽色、体态和比例参考 @Image 2。低机位中景,轻微手持呼吸感,主体始终位于画面中心。 0-3 秒:红冠鹤静立浅水,缓慢展开黑白双翼,翼风在水面推开涟漪;朝阳穿过薄雾形成自然体积光。 3-8 秒:红冠鹤轻跃,双足交替点水溅起水珠,随后轻盈落地收翼,抬头发出清亮长鸣。镜头小幅上下跟随跳跃轨迹。 保持自然景深,前景芦苇轻微虚化,主体清晰,背景晨雾与日出柔和虚化。仅保留流水、振翅与鹤鸣等自然环境声。 ``` Pattern: visual intent → asset identity/scene roles → short timeline → camera, focus, and sound convergence. ## Exact 30-second structure ```text 30 秒、16:9、单镜头情绪短片。清晨河畔,一位古装女子与镜头对面的爱人告别。电影写实质感,浅景深,柔和自然光,近景主观视角,轻微手持呼吸感,不快速切镜。 人物为 22 岁成年东亚女性:冷调白皙但保留真实微毛孔与自然皮肤纹理;细长含泪眼、柔和下颌线;黑发低髻,以素玉簪固定;白色交领轻纱衣。表演克制,重点表现视线转移、呼吸、嘴唇轻颤、吞咽和一滴泪自然滑落。 0-3 秒【追问】:直视镜头,尚未落泪,轻声问:“你真的要走吗?” 3-10 秒【接受】:视线缓慢移开,深呼吸,将未说出口的话咽下。 10-20 秒【不舍】:重新看向镜头,眼眶逐渐湿润,保持克制,不夸张哭泣。 20-27 秒【告别】:嘴角勉强露出温柔微笑,一滴泪在此阶段才滑落。 27-30 秒【余韵】:她轻轻点头,镜头保持近景,停在含泪但平静的目光上。 全程锁定人物身份、发型、服装、河岸机位和清晨光向;无额外台词、无字幕、无 BGM、无夸张肢体动作、泪水不得提前出现。 ``` Pattern: global world/character/performance lock → complete continuous timeline → timing-sensitive prohibitions. ## Long Video ```text 总时长 60 秒,16:9,电影写实、复古治愈风格。雨日下午的复古咖啡馆,固定机位记录木桌上的热咖啡和玻璃窗外雨景。 0-20 秒【安静开场】:镜头固定,咖啡热气缓慢上升;雨滴落在格纹玻璃上并汇聚下滑。无人进入画面,不切镜。 20-40 秒【细节推进】:镜头极慢推近咖啡杯;一片金色落叶飘过并贴在窗外玻璃上,焦点仍停在杯沿和热气。 40-60 秒【焦点转移】:焦点从前景咖啡缓慢转向窗外雨景,街灯在雨水中散成暖色光斑,画面随雨声逐渐变暗结束。 全程保持室内暖黄光与室外冷雨色的冷暖对比,只使用真实雨声和轻微环境声;无人物、无文字、无 BGM、无闪烁或突兀转场。 ``` Pattern: total duration and ratio → broad readable acts → one controlled evolution → sound and stability convergence. ## Video extension Natural continuation: ```text 参考 @Video 1 的原始内容,向后续写 5 秒。原视频内容保持不变,本提示仅作用于新增 5 秒。肌肉战士继续位于同一原始森林空间,承接边界帧中的步伐、身体朝向、镜头前移速度、斑驳日光和环境声;他突然察觉异常,停下脚步,目光变得锐利,顺势转入防御姿态并盯住前方。动作和镜头衔接自然流畅,不硬切,不重置站位,不让道具或角色凭空出现。 ``` Transition continuation: ```text 参考 @Video 1,向后续写 10 秒并继续故事;原视频保持不变。连接点使用遮挡转场:近景中的战士保持防御姿态,一片正常大小的落叶随风接近并完全遮住镜头;镜头穿过叶片后自然拉开,揭示 @Image 2 的年轻游侠倚树举弓瞄准。保持遮挡方向、镜头运动和风向连续,从战士近景自然切换为游侠中景;禁止硬切和物体凭空出现。 ``` ## Video edit Unmarked edit: ```text 在整段 0-8 秒内,将视频中的鲸鱼替换为白色东方龙;保持原有动作轨迹、主体尺度、镜头运动、场景、光影和其他元素不变。移除 BGM,保留其余指定现场声。 ``` Marked edit: ```text 仅修改红色框内、红色箭头所指的咖啡杯:在 5-8 秒将它逐渐替换为一束红玫瑰。框外人物、桌面其他物品、背景、机位、光影、动作和声音保持不变;玫瑰需要符合原视频透视、遮挡、接触阴影和运动稳定性。 ``` ## Clay Renderer Coarse model: ```text 参考 @Video 1 的相机轨迹、人物动作和物理光影变化,将白色人形模型完整映射为 @Image 1 的成年精灵女王;不要继承灰模材质、视口背景或辅助线。人物保持银白长发、白金刺绣长袍和水晶法杖,身份与服装全程一致。 场景渲染为精灵大厅,保持 @Video 1 的机位和空间调度。女王按灰模动作自然举起法杖,长发与宽大裙摆随气流产生连贯摆动;原视频光效转译为从地面升起的圣洁光柱,与法杖金色光芒呼应。 最终为电影级写实 CG,皮肤、金属刺绣和丝绸材质清晰,动作自然,无穿模、变脸、灰模质感、坐标轴或轨迹线。 ``` Fine model: ```text 将 @Video 1 的完整白模动画渲染为最终成片。0-7 秒保持写实海洋夜景:深灰蓝风暴海面持续起伏,红白条纹灯塔按白模动画随巨浪俯仰;使用低饱和电影氛围、HDR 光照、体积雾、湿润高光、真实反射和动态阴影。 风雨、浪花和雾气保持统一风向与物理逻辑;相机、构图、灯塔位置和运动轨迹严格保留。仅保留海浪、狂风、雨声与偶发低沉雷声,无 BGM、无旁白、无视口网格、轨迹线、坐标轴或相机锥体。 ``` ## Seamless transition ```text 无缝连接 @Video 1 与 @Video 2,但不要修改两段原视频的既有内容。@Video 1 的镜头继续快速推向水洼中的圆形蓝紫霓虹倒影;圆形倒影占满画面后,自然过渡到 @Video 2 的金色太阳倒影。持续推镜过程中,雨夜水洼纹理由外向内逐渐变成森林湖面的水波,蓝紫色霓虹同步转为金色阳光,环境随之完成变化。 连接点锁定中央圆形倒影的形状、屏幕位置、大小、推近方向和速度;镜头运动连续,无黑帧、跳帧、硬切、闪烁、文字、标志或字幕。 ``` ## Multi-grid storyboard ```text @Image 1 是包含 4 个连续镜头的完整分镜脚本,每一格代表一个完整镜头,而不是独立人物或场景。成年女性角色身份与服装参考 @Image 2;橘白小猫的品种、毛色、服装和比例参考 @Image 3,二者全程保持一致。 镜头 1(0-4 秒):黄昏住宅区,低机位中景缓慢推近,小猫跳上垃圾桶寻找食物。 镜头 2(4-8 秒):切至近景,小猫扒住桶沿翻找,耳朵轻动,突出饥饿和无助。 镜头 3(8-13 秒):固定中景,女性在门口发现小猫并放下食物,小猫保持警惕地接近。 镜头 4(13-15 秒):镜头缓慢拉远,小猫跟随女性走进暖光中的家门,停在安静庭院上结束。 采用温暖治愈的 3D 动画电影 CG 风格,保持时间、空间、人物位置、光向和动作连续;禁止变脸、比例漂移、毛色变化、服装变化以及新增无关人物或道具。 ``` ## Specialist snippets BGM removal: ```text 移除 @Video 1 的背景音乐,仅保留清晰人声和真实室内环境声;原画面、原有字幕、人物动作和镜头保持不变。 ``` Timbre binding: ```text 角色 A 在 6-9 秒说“我找到你了”,音色参考 @Audio 1,保留低沉音色、缓慢语速和压低呼吸的悬疑感;不要继承参考音频中的背景噪声或其他说话人。 ``` Multiple-person lock: ```text 新娘身份参考 @Image 1,新郎身份参考 @Image 2,三位伴娘统一参考 @Image 3,三位伴郎统一参考 @Image 4。身份、数量、服装和主次关系全程清晰;禁止新郎新娘混脸、伴郎伴娘服装互换、重复人物和额外肢体。 ``` Creative transfer: ```text 参考 @Video 1 的手机滑动动作、镜头节奏和转场逻辑,将其转译为围绕 @Image 1-5 内容展开的创意短片;保持目标图片的核心主体、空间结构和视觉风格,不继承参考视频中的人物身份、背景文字或标志。 ``` Perspective reconstruction: ```text 将 @Video 1 的多机位切换改为固定正面机位,人物身份、动作时间、背景空间和原声保持不变;保持稳定中景,不产生新的侧面或俯拍镜头。 ``` Green-screen composite: ```text 提取 @Video 2 的绿幕前景并合成到 @Video 1 的城市道路。当前景从镜头旁经过时,匹配道路透视、主体尺度、地面接触、阴影、运动模糊和相机下俯角;移除绿色溢色、边缘光晕、漂浮感和逐帧闪烁,原城市镜头的运动和空间保持不变。 ```references/prompt-blueprints.md9.9 KBd438b7eba79bupstream ↗
# Seedance 2.5 prompt blueprints ## Contents - Shared assembly order - Basic multimodal - Exact 30-second timestamp video - Long Video - Video extension - Video edit - Clay Renderer - Seamless transition - Multi-grid storyboard - Transition vocabulary Use one primary blueprint. Add only relevant specialist instructions from `multimodal-patterns.md`. ## Shared assembly order Prefer this semantic order inside the final copy-ready prompt: 1. Output declaration: mode intent, duration, ratio, resolution when relevant. 2. Asset declaration: what each referenced asset controls and what it must not contribute. 3. One-sentence film intent: subject + place + event + tone/style + special camera idea. 4. Global setting: environment, physical texture, visual style, camera family, subject design, core performance. 5. Chronological plot or edit operation. 6. When needed, request-specific performance, physical causality, live coverage, or cutting rules. 7. Audio: dialogue, sound effects, ambience, timbre, music, subtitle policy. 8. Global continuity and prohibited changes. Do not emit bracketed template labels in the final prompt unless they improve a long or technical prompt. ## Basic multimodal Use for a new 4-30 second video that does not require another primary mode. Formula: `asset roles + one-sentence intent + chronological visual/action description + camera and sound + global continuity/prohibitions` For a simple single-beat video, prose may be better than timestamps. Include timestamps only when the user requests timing or the action sequence needs disambiguation. Match control depth to the request. If the user supplies a detailed layered structure or asks for precise acting, physics, live coverage, or shot/reverse-shot direction, retain comparable hierarchy and granularity using `realistic-direction-patterns.md`; do not flatten it into the compact formula. ## Exact 30-second timestamp video Use for an exact 30-second video with deliberate pacing. Build: 1. `Reference layer`: bind identity, scene, composition, movement, timbre, or music sources. 2. `Global setting`: environment and physical texture; visual style and light; camera language; subject design; performance core; request-specific prohibitions. 3. `Timestamp storyboard`: divide the complete `0-30s` interval into readable phases. 4. `Global convergence`: restate continuity, ending state, audio, subtitle, and BGM policy. Each time segment should include: `time range + dramatic function + visible action/expression + camera/framing + sound or dialogue + intended end state` Use emotional interpretation only when it translates into visible acting, gaze, breath, posture, vocal delivery, or camera behavior. ## Long Video Use for a new 30-180 second video. Formula: `global parameters + asset roles + one-sentence overview + act/scene progression + global convergence` - State total duration and ratio at the beginning. - Prefer acts or scenes over excessive second-by-second microdirection. - Give each scene a narrative job, approximate or exact time range, location, action, camera strategy, and transition. - Maintain an explicit state chain across scenes: subject location, wardrobe, held props, time/weather, and unresolved goal. - Use exact timestamps only for beats that require synchronization. - End with a resolved visual state rather than a list of events. ## Video extension First normalize direction: - `prepend-before-source`: insert the new interval before the original. The new interval must end at the source's first-frame state and hand camera, motion, lighting, space, and audio into the original opening. - `append-after-source`: add the new interval after the original. The new interval must begin from the source's last-frame state and continue camera, motion, lighting, space, and audio from the original ending. - Treat `原片之前 / 补前情 / 前传 / 衔接首帧` as prepend evidence. Treat `原片之后 / 接着结尾 / 然后 / 继续故事` as append evidence. - Treat bare `向前续写 / 前向续写 / 往前延长 / 向后续写 / 后向续写` as linguistically ambiguous. Do not resolve these phrases from plot progression, opening-state descriptions, or common-language assumptions. Unless the user explicitly places the new interval before or after the original or names the first/last-frame connection, ask `新增片段放在原片之前,还是原片之后?` before drafting. Required opening after normalization: `参考 @Video N 的原始内容,在原片之前插入/在原片之后追加 X 秒。原视频区间保持不变,以下提示仅作用于新增区间。` Then specify: 1. Relevant source boundary state: for prepend, the required destination state from the first frame; for append, the inherited start state from the last frame. 2. New action or scene only for the added interval, arranged in chronological output order. 3. Connection behavior: natural continuation into the original opening for prepend, or out of the original ending for append; use a named transition only when requested. 4. End state of the added interval: match the original first frame for prepend; define the new final state for append. 5. Constraints: smooth action and camera connection; no unexplained reset, hard cut, sudden prop, identity change, lighting jump, or audio discontinuity. For a transition extension, use: `transition guide + source-boundary state + destination scene + connection logic + framing change` ## Video edit Choose the operation subtype: - `Smart Edit`: text-targeted change to a generated video. - `Edit with marks`: change located by a box, brush, arrow, line, text, or landmark. - `Edit Video`: equivalent local editing applied to an uploaded source video. Formula without marks: `exact target + add/remove/replace/change instruction + effective time + protected elements` Formula with marks: `annotation tool and inside/outside/direction + exact target + add/remove/replace/change instruction + effective time + protected elements` Use an explicit A-to-B change. Example structure: `在 0-8 秒内,将画面中央人物的蓝色牛仔裤替换为深黑色西装裤;人物身份、上衣、动作、背景、机位和原声保持不变。` For removal, require physically plausible background completion consistent with original perspective, motion, texture, light, and occlusion. ## Clay Renderer ### Coarse white model Use when simple geometry primarily controls dynamic structure. Formula: `reference declaration + model-to-final-subject mapping + chronological action/camera plot + scene treatment + global convergence` - Name exactly which motion trajectory, camera move, position, scale relation, lighting change, cut rhythm, or sound timing to borrow. - Map every distinguishable model by color, shape, or position to a final character, object, or prop. - Describe complete limb or wing action chains if present; incomplete actions risk stiffness. - Do not preserve primitive geometry, gray material, guide overlays, or viewport texture. ### Fine white model Use when geometry, scene, and animation are complete and need final rendering. Formula: `render command + timed rendering description + scene treatment + global convergence` - Start with `将 @Video N 的白模动画渲染为最终成片。` - Describe environment, tone, material, surface response, light, atmosphere, and timed scene changes. - Preserve camera, composition, animation, relative positions, and intended occlusion unless the user asks to alter them. - Remove trajectory lines, coordinate axes, camera cones, viewport grids, and other production overlays. ## Seamless transition Required opening: `无缝连接 @Video 1 与 @Video 2,但不要修改两段原视频的既有内容。` Define: 1. Exit action or visual trigger from Video 1. 2. Full-screen bridge state: occlusion, matching object, blur, particles, light, surface, or camera passage. 3. Entry action into Video 2. 4. Connection locks: shape, screen position, scale, direction, speed, motion trend, camera vector, color/material transformation, and sound bridge. 5. Prohibitions: no black frame unless intentional, frame skip, hard cut, flicker, text, logo, subtitle, or alteration of the source clips. ## Multi-grid storyboard Build: 1. Declare `@Image 1` as the complete ordered storyboard and state that each panel is one complete shot, not an independent character or scene reference. 2. Map separate character, prop, and environment references. 3. State the story spine and ending. 4. For every panel/shot, provide exact or proportional time, composition, camera height/framing, subject action, spatial relationship, and transition/end state. 5. Close with global style, physical behavior, continuity, and prohibited changes. Prefer simple line or stick-figure boards for geometry; the prompt must supply appearance, environment, movement, and rendering detail. ## Transition vocabulary Use transitions only when narratively and physically motivated. | Transition | Best use | Control to state | |---|---|---| | Natural shot switch | Conventional narrative | breathing room, matching action/gaze, no abrupt timing | | Fade in/out | Opening, ending, passage of time | black/white state and duration | | Dissolve | Memory, dream, lyrical time passage | overlap duration and matching composition | | White/black flash | Impact or beat sync | trigger sound/action and exact flash moment | | Wipe | Clear location change, retro grammar | wipe direction, speed, covering edge | | Occlusion/mask | Spatial passage | occluding object, full-cover state, reveal direction | | Match object | Montage | shared shape, position, scale, color, motion | | Whip/action transition | Action, travel, speed | pan direction, force, blur, entry direction | | Motion relay | Outfit/location change | continuous body action and landing pose | | Zoom in/out | Macro-to-micro or nested world | target detail, full-frame moment, destination reveal | | Ink diffusion | Eastern poetic transition | ink trigger, spread direction, reveal timing | Never add a transition merely to make a prompt sound cinematic.references/realistic-direction-patterns.md9.1 KB04611591d446upstream ↗
# Realistic direction patterns ## Contents - Control depth - Example-transfer contract - Natural human performance - Dialogue and emotional conflict - Relationship-drama camera grammar - Sports and variety live coverage - Physical causality - Contradiction audit Use only the sections relevant to the request. These are directing heuristics, not extra Seedance modes. ## Control depth Match prompt depth to the user's intent and source material. - **Compact:** one subject, one beat, usually under 10 seconds. Use intent, action, camera, sound, and essential locks. - **Sequenced:** a 10-30 second scene with several state changes. Use global setup plus readable chronological beats. - **High-control:** use when the user asks for detail, supplies a layered example, needs exact performance/physics, or requires live multicamera grammar. Use asset locks, global world, timed execution, performance or physics, audio, camera rules, and final invariants. Do not mistake brevity for stability. Reduce conflicting or impossible instructions first; preserve detail that controls identity, causality, timing, camera purpose, or emotional state. ## Example-transfer contract When a user supplies an example, separate these dimensions: 1. hierarchy and section order; 2. detail density and timestamp granularity; 3. story events and ending; 4. camera POV, shot family, and cutting grammar; 5. performance logic and emotional arc; 6. visual style, era, light, and texture; 7. asset-reference scope; 8. dialogue, sound, and subtitle policy. Transfer only explicit dimensions. `参考这个结构`, `按这个结构写`, or `只参考提示词结构` means inherit hierarchy, control depth, and comparable granularity only. Do not inherit first-person POV, one-take form, wardrobe, plot, dialogue, props, setting, or ending unless requested. When revising, propagate every story change into the timeline, physics/performance section, sound, camera rules, continuity, and final mandatory sequence. Remove stale constraints from the old version. ## Natural human performance Direct a motivated reaction chain: `stimulus → comprehension pause → first involuntary response → chosen response → residual emotion → next beat` - Put the character's objective, fear, defense, and contradiction before microgestures. - Use 1-3 visible cues per beat: gaze shift, breath catch, incomplete gesture, grip change, posture change, swallowed line, or delayed reply. - Let the previous emotion remain during the next beat. A smile fades, laughter decelerates, anger leaks into hurt, and resolve can coexist with affection. - Allow natural asymmetry, blinking, small focus drift, self-correction, interrupted gestures, and conversational overlap. - Describe anatomy only when it is visually necessary. Avoid sequential checklists for eyebrows, eyelids, nostrils, cheeks, and mouth; they produce mechanical acting. For conflicted emotion, show two impulses through behavior. Examples: a hand loosens from a suitcase handle and grips it again; a character almost steps closer, then restores distance; lips part to agree, then the gaze falls and the answer changes. ### Blush, tears, and laughter - Treat blush as optional. If needed, make it slight, gradual, uneven, and physiologically motivated. Prefer gaze avoidance, altered breath, posture, and speech rhythm over sudden uniform red cheeks. - Accumulate tears before release: breath changes, eyes moisten, blink timing changes, then one tear may cross the lid at the specified late beat. Do not start with full tears unless the scene starts after prior crying. - Build laughter from suppression or a trigger, then preserve residual breath and shoulder movement. Do not jump from neutral to exaggerated laughter and back to neutral. - Preserve real skin texture; prohibit beauty-filter blush, plastic skin, and emoji-like expression switching when realism matters. ## Dialogue and emotional conflict Budget dialogue for performance, not reading speed. As a rough Mandarin heuristic, allow about 3-5 Chinese characters per second for natural emotional delivery; a brief argument can reach 5-6, while hesitation, crying, or subtext needs more time. Count pauses, listening, interruption, breath, and action inside the interval. Build short conflict scenes around: 1. each character's immediate objective; 2. the surface issue; 3. the defense or misunderstanding; 4. the deeper reveal; 5. the listening reaction; 6. a choice or changed relationship state. Keep dialogue conversational: fragments, self-corrections, incomplete sentences, and repeated everyday phrasing are often more credible than polished slogans. Give the listener screen time. The decisive emotional beat may be the reaction after a line, not the line itself. For a breakup or reconciliation, avoid writing one character as emotionally finished unless requested. Show attachment and resolve together: instinctively defending the partner, reacting to a familiar promise, nearly yielding, then choosing based on accumulated history. ## Relationship-drama camera grammar First obey explicit form: POV versus third person, one take versus edited coverage, same-frame permissions, shot-scale limits, and aspect ratio. For conventional third-person drama: - establish location, prop, distance, and the active relationship state with a wide or medium-wide shot; - use clean singles or shot/reverse-shot as requested; - maintain the 180-degree axis, screen direction, eyeline height, light direction, and background geography; - cut for a reason: speaker emphasis, listener comprehension, a withheld reaction, a prop action, or a changed power balance; - enter closer framings as emotional pressure increases, but do not fill the frame with a face when prohibited; - hold after important lines so the reaction can develop before the next cut; - use J- or L-cuts only when they preserve spatial clarity and conversational flow. If `不要两人同框`, exclude shoulders, silhouettes, reflections, mirrors, foreground blur, and accidental background appearance of the other principal. An establishing shot can contain only one character plus environment unless the user explicitly permits both. For one-take scenes, timestamp headings are performance beats, not cuts. Change framing only through plausible camera or subject movement; do not use cut language, viewpoint jumps, or impossible lens changes. Do not infer POV or handheld style from an example described as a structural reference. ## Sports and variety live coverage Treat live coverage as action explanation, not generic cinematic motion. 1. Establish course geography, travel direction, obstacle order, finish state, and camera placement. 2. Assign each camera a purpose: crane/wide for geography, long lens for anticipation, side track for locomotion, low/water camera for scale and contact, ENG medium for recovery or reaction, finish camera for resolution. 3. Cut at action nodes while preserving progress, screen direction, body state, obstacle phase, and the subject's wetness/fatigue. 4. Show enough of the obstacle before impact so the viewer understands why contact occurs. 5. Make commentary, crowd, mechanism sounds, music, and graphics react after the visible event. Use a high-control layered structure when requested: `global live style → complete character lock → course continuity → exact timeline → physical feedback → performance arc → sound design → camera rules → final mandatory sequence` Do not add unplaceable orbiting cameras, cinematic teleportation, repeated actions, or a replay unless requested. Use fictional, minimal broadcast graphics; avoid real network branding and unreliable dense text. ## Physical causality For obstacles, falls, collisions, water, props, clothing, and other consequential action, write: `cause or mechanism → approach → contact → force direction → body/object response → recovery or outcome → next state` - No body reaction before contact. - No fall, splash, button activation, light, or sound before its trigger. - Preserve gravity, inertia, surface traction, object rotation, pendulum phase, water displacement, and contact shadows at the level visible in the shot. - Keep wardrobe, hair, wetness, fatigue, damage, held props, and location changes cumulative. - Reduce event count if the full causal chain cannot fit the interval. ## Contradiction audit Resolve these before drafting: - casual phone video versus extreme shallow depth of field or impossible stabilized motion; - fixed social distance versus face-filling close-up; - one take versus cut, reverse-shot, or instant viewpoint change; - third person versus camera owned by a character; - no two-shot versus shoulders, reflections, or silhouettes of both characters; - full-character reference versus redesigned wardrobe or accessories; - exact duration versus dialogue and actions that cannot fit; - natural acting versus muscle-by-muscle expression checklists; - no BGM versus requested score; - protected source footage versus a global restyle or edit; - changed story outcome versus stale physics, sound, camera, or final constraints from the old outcome. When a lower-priority style wish conflicts with an explicit camera, identity, duration, or edit lock, preserve the lock..github/workflows/validate.yml388 B57f9d71ad822upstream ↗
name: Validate Skill on: push: pull_request: permissions: contents: read jobs: validate: runs-on: ubuntu-latest steps: - name: Checkout uses: actions/checkout@v4 - name: Set up Python uses: actions/setup-python@v5 with: python-version: "3.11" - name: Validate repository run: python scripts/validate_skill.py.gitignore60 B166f904d3ce7upstream ↗
__pycache__/ *.py[cod] .DS_Store Thumbs.db .idea/ .vscode/agents/openai.yaml318 B39a3106bc2efupstream ↗
interface: display_name: "Seedance 2.5 视频导演" short_description: "生成真实表演、类型化镜头、视频脚本与 Seedance 2.5 提示词" default_prompt: "使用 $seedance-2-5-video-director 把我的视频创意或范例整理成真实可演的视频脚本、导演方案或可复制提示词。"LICENSE1.0 KBff7b45fe5a5eupstream ↗
MIT License Copyright (c) 2026 liyue-aigc Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.NOTICE.md718 B42b08ba58f4cupstream ↗
# Third-party notice This is an independent community project and is not affiliated with, endorsed by, or sponsored by ByteDance, Dreamina, 即梦, or Seedance. The files `references/capabilities-and-limits.md` and `references/official-examples.md` contain independently organized summaries, normalized structures, and shortened examples derived from the Dreamina Seedance 2.5 user guide identified in those files. They are included for interoperability and instructional reference. Platform names, trademarks, documentation, and service behavior remain the property of their respective owners. The runtime UI and current provider documentation take precedence over this repository when limits or features change.tests/cases.json4.9 KB0cd098ac77d9upstream ↗
[ { "id": "ordinary-new-conversation-skips-guide", "category": "onboarding", "request": "新开对话后使用 $seedance-2-5-video-director,生成一个10秒咖啡广告提示词。", "expected": [ "不展示安装或新手教程", "直接完成咖啡广告请求", "不把新对话误判为刚安装" ] }, { "id": "explicit-tutorial-request-shows-guide", "category": "onboarding", "request": "请显示 $seedance-2-5-video-director 的使用教程和示例。", "expected": [ "显示学习指导、最小输入模板和学习示例", "不声称每次新对话都必须重复", "不要求提供视频任务" ] }, { "id": "basic-person-reference-lock", "category": "basic-multimodal", "request": "用 @Image 1 的人物生成15秒街头行走视频,严格参考人物图片。", "expected": [ "选择basic-multimodal", "把@Image 1解释为完整可见人物参考", "保持原服装、鞋子、配饰和身体比例", "默认不继承参考图背景和姿势" ] }, { "id": "exact-thirty-water-variety", "category": "timestamp-30s", "request": "生成30秒、16:9水上障碍综艺,要求精确时间轴和中文直播镜头。", "expected": [ "选择timestamp-30s而不是Long Video", "时间段连续且总和等于30秒", "镜头解释赛道、接触和结果", "解说与观众在事件发生后反应" ] }, { "id": "long-video-scene-acts", "category": "long-video", "request": "生成90秒、16:9的长视频,讲一名女孩从车站走到海边的完整故事。", "expected": [ "选择long-video", "使用场景或幕结构而不是每秒微操", "维护地点、服装、道具和时间状态链", "结尾落在明确视觉状态" ] }, { "id": "extension-ambiguous-direction", "category": "video-extension", "request": "参考 @Video 1 向前续写5秒,让人物走进房间。", "expected": [ "识别向前续写的语言歧义", "只询问新增片段在原片之前还是之后", "方向未确认前不擅自起草提示词" ] }, { "id": "extension-append-boundary", "category": "video-extension", "request": "在 @Video 1 原片之后追加5秒,承接最后一帧让人物关门。", "expected": [ "规范化为append-after-source", "只描述新增5秒", "保护原视频", "继承最后一帧的动作、机位、光线、空间和声音" ] }, { "id": "marked-local-edit", "category": "video-edit", "request": "在 @Video 1 的5到8秒,只把红框内蓝色杯子替换成白色花瓶。", "expected": [ "选择带标记的视频编辑", "写清红框、目标、替换操作和时间", "保护框外人物、动作、背景、机位和声音", "要求透视、遮挡和接触阴影一致" ] }, { "id": "coarse-clay-renderer", "category": "clay-renderer", "request": "把 @Video 1 的粗白模动作映射为 @Image 1 的女王并渲染宫殿场景。", "expected": [ "选择coarse Clay Renderer", "映射白模动作、位置和相机", "保持@Image 1人物身份与服装", "不继承灰模材质、辅助线和视口背景" ] }, { "id": "two-video-seamless-bridge", "category": "seamless-transition", "request": "无缝连接 @Video 1 与 @Video 2,用一片树叶完全遮挡镜头完成转场。", "expected": [ "保护两段原视频不变", "定义树叶进入、全屏遮挡和离开", "锁定方向、速度、大小和相机连续性", "禁止黑帧、跳帧和硬切" ] }, { "id": "ordered-grid-storyboard", "category": "multi-grid-storyboard", "request": "@Image 1 是四格连续分镜,@Image 2 是主角,请生成15秒视频。", "expected": [ "把@Image 1每一格解释为有顺序的完整镜头", "把@Image 2绑定为主角身份", "为每格分配时间、构图、动作和结束状态", "保持跨格空间与人物连续" ] }, { "id": "diagnose-camera-contradictions", "category": "diagnosis", "request": "只诊断:一镜到底、固定机位,同时频繁正反打;人物始终离镜头两米,最后脸部填满画面。", "expected": [ "指出一镜到底与正反打冲突", "指出固定距离与脸部填满画面冲突", "不静默重写完整故事" ] }, { "id": "script-only-breakup", "category": "basic-multimodal", "request": "只要25秒第三人称分手戏脚本,不要两人同框,女主既舍不得又决定离开。", "expected": [ "只输出可演脚本", "使用单人近景或正反打且不出现双人同框", "保留听见、停顿、反应和残留情绪", "对白和动作适配25秒" ] } ]
Plain-text copy, no login: /skill-bundles/liyue-aigc-seedance-2-5-video-director.md