Content & Media Creation · prime-skills/runcomfy-agent-skills

seedance-v2

Name: seedance-v2
Author: prime-skills/runcomfy-agent-skills

Generate cinematic short-form video with ByteDance Seedance 2.0 Pro on RunComfy. Documents Seedance 2.0 Pro's strengths (multi-modal references — up to 9 images, 3 videos, 3 audio — synchronized in-pass audio with natural lip-sync, cinematic motion refinement), the 4–15s duration schema, and when to route to HappyHorse 1.0 / Wan 2.7 / Kling instead. Calls `runcomfy run bytedance/seedance-v2/pro` through the local RunComfy CLI. Triggers on "seedance", "seedance 2", "seedance v2", "seedance pro", "bytedance video", or any explicit ask to generate video with this model.

L4 High-Privilege · Graded auth & full auditV1 Commit-PinnedExecute CommandNetwork AccessWrite FileDual-pass Audit

Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

Implemented as "zero-code + external CLI delegation": `seedance-v2/` contains only a single `SKILL.md` with no scripts; at runtime, the host agent executes `runcomfy run bytedance/seedance-v2/pro --input '<JSON>' --output-dir <path>` via Bash. Video generation and audio synthesis occur entirely on the RunComfy cloud, with the skill solely responsible for translating intent into proper JSON inputs and prompts.

seedance-v2/SKILL.md

The skill invokes `runcomfy run bytedance/seedance-v2/pro` with a JSON body matching the schema.

`git ls-tree HEAD` confirms that `seedance-v2/` contains only `SKILL.md`; all 31 skills across the same repository share this identical pattern.

The core differentiator is multi-modal reference conditioning: 0–9 images, 0–3 videos, and 0–3 audio tracks can be fed simultaneously, working in tandem with `generate_audio` to produce synchronized speech, sound effects, and background music in a single pass—making it the only model in the family supporting "multi-type references + in-pass speech".

seedance-v2/SKILL.md

| `video_url` | array | no | `[]` | 0–3 clips (MP4/MOV), 2–15s each. |

Each of the three reference types enforces distinct format and duration constraints (both video and audio require 2–15s, audio < 15MB, images support JPEG/PNG/WebP/BMP/TIFF/GIF); these specifications are reiterated across schema, Prompting, and Limitations sections, rejecting out-of-bound inputs directly.

The schema comprises 10 fields, including 7 aspect ratio values (defaulting to `adaptive`), durations of 4–15 seconds (defaulting to 5), resolutions of 480p/720p (defaulting to 720p), `generate_audio` enabled by default, and `seed`. The aspect ratio enumerations (including 21:9 and `adaptive`) along with the 4s minimum duration floor differ from sibling models and reflect the model's native specifications.

seedance-v2/SKILL.md

| `aspect_ratio` | enum | no | `adaptive` | `adaptive`, `16:9`, `9:16`, `4:3`, `3:4`, `1:1`, `21:9`. |

`adaptive` as a default value is unique within the family (`wan-2-7` and `happyhorse` default to 16:9); Limitations explicitly state "Resolution ceiling 720p on the playground variant", candidly documenting the resolution compromise made for its "cinematic" positioning.

The most crucial operational discipline is "division of labor between image and text": stable identity (face, clothing, branding, logo) goes into `image_url`, while evolving narrative (action, atmosphere, lighting, camera angle) goes into `prompt`; attempting to describe facial features verbally wastes tokens and induces drift.

seedance-v2/SKILL.md

**Image vs text division.** This is the single most important rule. Stable identity (face, costume, brand mark, logo) → put in `image_url`. Evolving narrative (action, mood, lighting, camera) → put in `prompt`. Trying to verbally describe a face in detail wastes tokens and produces drift.

This turns "multi-reference" from a parameter issue into an explicit methodology—the documentation even names the antipattern directly: "Trying to describe stable identity verbally → use `image_url` instead."

Audio direction and lip-sync are jointly controlled by `generate_audio` and prompts: synchronized speech, sound effects, and background music are generated in the same pass by default; the documentation requires explicit mood and atmosphere descriptors (e.g., "warm friendly conversational / gentle cafe chatter, no foreground music") and acknowledges that lip-sync quality depends on prompt clarity.

seedance-v2/SKILL.md

| `generate_audio` | bool | no | true | In-pass synchronized speech / SFX / music. |

Limitations explicitly state "Lip-sync quality — depends on prompt clarity; not guaranteed perfect under all conditions.", establishing realistic expectations upfront rather than serving as an afterthought disclaimer.

The routing table clearly demarcates responsibilities across sibling video models: route to Wan 2.7 (`audio_url`) for audio-driven lip-sync, HappyHorse 1.0 for top-tier benchmark visual quality, Kling for motion editing of existing footage, and LTX 2 for rapid iteration; if the user explicitly mentions Seedance, route here unconditionally.

seedance-v2/SKILL.md

| Audio-driven lip-sync from your own track | Wan 2.7 (`audio_url`) |

The page states: "If the user said \"Seedance\" / \"Seedance 2\" / \"ByteDance video\" explicitly, route here regardless."; among recommended targets, Wan 2.7 and HappyHorse 1.0 have corresponding skill directories in-repo, whereas "Kling Video O1" and "LTX 2" do not (see second_pass).

Execution chain and boundary limitations: The CLI submits requests to the Model API at `/v1/models/bytedance/seedance-v2/pro`, polls for results, and downloads whitelisted artifacts; the documentation acknowledges a 720p ceiling, absence of `@` character binding syntax, and 422 errors for durations >15s requiring chunking.

seedance-v2/SKILL.md

The CLI submits, polls, fetches the result, downloads `*.runcomfy.net`/`*.runcomfy.com` URLs into `--output-dir`.

Noting "no `@` syntax" explicitly contrasts its capabilities against competitors (which use `@` for character binding), providing a helpful negative constraint list for users.

2Core Capabilities

01Multi-modal reference conditioning: 0–9 images, 0–3 videos, and 0–3 audio tracks can be fed simultaneously

02In-pass generation of synchronized speech / sound effects / background music (`generate_audio` enabled by default) with natural lip-sync

03Image/text division-of-labor discipline (identity in reference images, narrative in prompt) to prevent drift from verbal face descriptions

04Aspect ratio, duration, and resolution control (7 aspect ratios including 21:9 and default `adaptive`, 4–15 seconds, 480p/720p)

05Strict reference asset specification discipline (video/audio 2–15s, audio < 15MB; out-of-bounds inputs rejected)

06Audio direction prompting guidelines (explicitly specify tone and atmosphere, e.g., warm friendly conversational / no foreground music)

07`seed` reproducibility and variant testing over a fixed schema

08Sibling model routing decisions (audio-driven lip-sync → Wan 2.7, benchmark quality → HappyHorse 1.0, motion editing → Kling, rapid iteration → LTX 2)

3External Dependencies

Kind	Dependency / Endpoint
cli	runcomfy（RunComfy 官方 CLI）
package	@runcomfy/cli（npm 全局包）
package	npx skills（skills.sh 安装器）
api	RunComfy Model API（ByteDance Seedance 2.0 Pro）
network	RunComfy 产物下载白名单
network	用户传入的参考图/视频/音频 URL（最多 9+3+3，由 RunComfy 服务器抓取，非本机）
network	RunComfy 站点/模型页/文档（文档引用，非运行时调用）
network	GitHub（技能源码链接）

4Recommended Governance & Evidence

Suggested Controls for Safe Operation

For Agent Runtime

Graded/per-use authorization, sandbox recommended

For Administrators

Approval workflow, rollback support, full audit

Evidence & Triggers (File Paths & Excerpts)

scripts/: Local script or command execution: (No scripts within the skill itself) Executed by the host agent via Bash using an external CLI: `runcomfy run <vendor>/<model>/<endp
网络端点: Declared external network access: POST https://model-api.runcomfy.net/v1/models/bytedance/seedance-v2/pro (via r
本地工作目录: File I/O and artifact output: Generated outputs are downloaded and written to `--output-dir` by the external CLI (documentation states downloads abort if a single file exceeds 2 GiB to prevent disk saturation)

Audit Notes: Based on the unified rating criteria (where `image-edit` in the same repo was rated Orange): At runtime, it drives a third-party npm package/CLI `@runcomfy/cli`, which reads user RunComfy credentials (`~/.config/runcomfy/token.json` or `RUNCOMFY_TOKEN`) on each invocation—triggering the "invokes third-party CLI/remote package touching credentials/quotas" tier. Assertions regarding CLI implementation, Model API behavior, 0600 token permissions, the 2 GiB download ceiling, and exit code semantics reside outside the repository and lack verifiable source code in this pin, warranting an Orange rating under the principle of "escalate when uncertain". Furthermore, this skill exhibits the broadest outbound asset footprint (up to 9 images + 3 videos + 3 audio tracks) and synthesizes voice in-pass—meaning confidentiality risks for unreleased assets and authentic human voice samples cannot be controlled locally; video generation also entails high-cost, pay-per-call inference. The skill itself is pure Markdown with no scripts, obfuscation, or hidden endpoints, thus avoiding the Red tier.

5Second-Pass Verification

[ok] skill 目录内容 = 仅 SKILL.md — git ls-tree HEAD 与 ls -a 双向核实；仓库根仅 LICENSE + README.md + 31 个技能目录。
[ok] 全部外部依赖条目回查调用点真实存在 — runcomfy CLI 调用点 3 处（默认/带角色参考/多模态示例）、npm 安装 1 处、npx skills 1 处、model-api 端点 1 处、白名单与外部素材抓取各 1 处、文档/模型页/GitHub 链接多处，均实读命中。
[discrepancy] 路由建议的目标模型在本仓有对应技能 — 'Motion editing on existing footage | Kling Video O1' 与 'Ultra-fast iteration | LTX 2' 在仓库内没有同名技能目录（全仓无 ltx-* 目录；kling-3-0/SKILL.md 自称 'Kling 3.0 (also called Kling V3.0)'，未使用 'Kling Video O1' 这一名称）。Wan 2.7 与 HappyHorse 1.0 两条路由目标在仓内有同名目录，属有效引用；Kling/LTX 两条为悬空引用。
[ok] 『cinematic 视频』定位与 720p 分辨率上限的一致性 — description 称 cinematic short-form，schema 与 Limitations 均明示 resolution 仅 480p/720p（'Resolution ceiling 720p on the playground variant'）。定位与规格存在张力，但文档主动披露上限，未夸大为 1080p/4K，判为一致。
[unlocatable] Security & Privacy 的技术断言可在本仓证实 — 0600 token 权限、2 GiB 下载上限、无 shell 展开、下载白名单、退出码语义均为仓外 @runcomfy/cli 行为；仓内无 CLI 源码或版本 pin，只能作为文档自述引用（同仓 runcomfy-cli/SKILL.md 为同源转述，非独立证据）。
[discrepancy] 安装一行与 GitHub 链接指向旧仓 agentspace-so/runcomfy-skills — 安装行 `npx skills add agentspace-so/runcomfy-skills --skill seedance-v2 -g` 与 GitHub 链接 github.com/agentspace-so/runcomfy-skills 指向旧仓；本仓 pin 在 prime-skills/runcomfy-agent-skills（README 用 runcomfy-agent-skills）。gh api：旧仓仍存在（5★、pushed 2026-04-29、含 seedance-v2 目录）；prime-skills/runcomfy-agent-skills 与 agentspace-so/runcomfy-agent-skills 为同一仓（相同 stats，full_name=prime-skills/runcomfy-agent-skills）。按文档安装可能装到旧版。
[discrepancy] 音/视频合成与声音权的风险披露 — 本技能会在同趟合成语音（generate_audio 默认 true），且接受 0–3 段参考音频引导音色；Security & Privacy 段仅泛泛提示 'image-based prompt injection'，未就『合成语音可用于伪造口播』或『参考音频可近似真人声音（声音权）』给出任何提示（对比：happyhorse 明写音频为 in-pass only、不接受外部音轨）。属安全提示覆盖不全。
[ok] prompt 长度规格是否自洽 — schema 写 'CN ≤ 500 chars OR EN ≤ 1000 words'，正文未另立口径（本族其它技能给的是字符/token 上限），无冲突。

6Verdict & Scenarios

Broadest multi-modal reference coverage (9 images + 3 videos + 3 audio tracks) coupled with an explicit division-of-labor methodology: identity in images, narrative in text, avoiding drift caused by verbal face descriptions.

In-pass synchronized audio eliminates standalone TTS workflows: talking-head ads and dialogue scenes generate speech and lip-sync in a single pass, enabled by default with prompt-driven tone and ambiance controls.

Strict reference asset specs reiterated across three sections (2–15s, <15MB, allowed formats list), explaining upfront that out-of-bound inputs are rejected to prevent wasted calls.

Transparent capability boundaries: 720p ceiling, lack of `@` character binding syntax, lip-sync quality contingent on prompt clarity, and mandatory chunking for >15s are all documented under Limitations.

Zero-code, single-file implementation, allowing the full directory to be audited before deployment.

Best for: Best suited for: Brand narrative shorts with character/scene reference assets (multiple character stills, locked-shot video, reference vocal timbre); conversational ads requiring single-pass generation with voiceover and sound effects (bypassing standalone TTS and lip-sync pipelines); engineering pipelines requiring fixed `seed` values for reproducible variant testing; and multilingual narratives (where image/text division allows translation by editing text only). Prerequisite: `runcomfy` CLI installed and authenticated, with acceptance of the 720p ceiling.

Not for: Not suited for: 1080p/4K delivery requirements (switch to `happyhorse-1-0` at 720p/1080p or alternatives); precise lip-sync driven by custom audio tracks (switch to `wan-2-7` via `audio_url`); single continuous shots >15s (requires chunking); `@` syntax character binding (unsupported by this model); or compliance environments where assets cannot leave the internal network or auditable execution code is strictly required.

Install Agent-Ready

① Verified Mirror 2026-09-15

Option A · Direct DownloadDownload seedance-v2.tar.gz

sha256: 0cbe7963b5552282…

Option B · Copy Guide for Agent

install.json

Agents automatically download and verify mirrors from this JSON guide.

② Upstream GitHub

GitHub Repo ↗

Pinned to commit fca19ae084.