Hubbestaiskill.com

Home / All Skills / Content & Media Creation / seedance-2-5-reference-to-video

Content & Media Creation · genmedia-labs/skills

seedance-2-5-reference-to-video

Generate reference-guided 1080p video with ByteDance Seedance 2.5 Reference to Video on RunComfy via the `runcomfy` CLI. Feed up to 9 reference images, 1-3 reference video clips, and 3 reference audio files into one call and get a 4-30 second 1080p clip with native synchronized audio, identity and style locked to your references. Documents the full input schema (images / videos / audios / aspect_ratio / duration / generate_audio), the counted-seconds billing model ($0.53 per second of reference video duration plus output duration), the 480p draft-then-deliver workflow, and when to route to Seedance 2.5 text-to-video, image-to-video, or Seedance 2.0 Pro instead. Calls `runcomfy run bytedance/seedance-2.5/reference-to-video/1080p`. Triggers on "seedance 2.5", "seedance 2.5 reference to video", "reference to video", "reference-to-video", "seedance 1080p", "ByteDance Seedance 2.5", "consistent character video", "style-locked video", or any explicit ask to generate video from reference images and clips.

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

Author genmedia-labsGitHub genmedia-labs/skills ↗Stars 7License MIT（仓库根 LICENSE：Copyright (c) 2026 RunComfy / agentspace-so；本 skill frontmatter 亦声明 license: MIT）Commit c163a9fd3a

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

Zero-code, single file: The entire skill consists of a single-page model document and calling recipe. The repository contains no scripts, no template files, and no binaries (all 36 directories in the repo contain only Markdown, with LICENSE being the only non-md file). Capabilities are not implemented within the skill; instead, it clearly documents the correct invocation method for the model endpoint and delegates execution to the official CLI.

seedance-2-5-reference-to-video/SKILL.md

allowed-tools: Bash(runcomfy *)

The frontmatter declares only one tool surface: the runcomfy subcommand. Finding non-.md files across the repo reveals only LICENSE, confirming no hidden scripts or data.

The end-to-end call chain is documented and auditable: Skill constructs JSON body → executes `runcomfy run <model_id>` → CLI POSTs to the corresponding model path on model-api.runcomfy.net → polls status → retrieves results → downloads output URLs from *.runcomfy.net / *.runcomfy.com to --output-dir.

seedance-2-5-reference-to-video/SKILL.md

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

The 'How it works' section details the full path: POST to `https://model-api.runcomfy.net/v1/models/bytedance/seedance-2.5/reference-to-video/1080p`. The description of the same process (Submit → poll → Download) in runcomfy-cli/SKILL.md within the same repository cross-verifies this.

The billing model itself serves as the selection pivot: This endpoint charges based on counted seconds (counted seconds = reference video duration + output duration) at $0.53/counted second, while image and audio references do not count toward duration—directly driving the operational decisions of 'whether to include reference video' and 'whether to trim reference clips'.

seedance-2-5-reference-to-video/SKILL.md

Billing is **$0.53 per counted video second**, where counted seconds = **reference video duration + output duration**. Image and audio references are not billed as duration.

The documentation translates this into actionable steps: 'trim reference clips before uploading' and 'use the 480p tier for reference selection'.

A two-stage workflow (480p draft → 1080p delivery) is advocated as the proper usage pattern: first run 3-5 reference combination variants using the 480p variant of the same endpoint family to evaluate only identity retention, shot matching, and audio alignment (ignoring resolution), then run the winning body verbatim at 1080p.

seedance-2-5-reference-to-video/SKILL.md

Re-run the winning body verbatim against `.../reference-to-video/1080p`, raising `duration` only once the beat is right.

The rationale is provided immediately after: $0.12/counted second for 480p vs. $0.53 for 1080p, making drafting costs roughly an order of magnitude lower than the final render.

Version migration pitfalls are explicitly highlighted: Seedance 2.0 Pro used image_url/video_url/audio_url, whereas 2.5 switched to images/videos/audios arrays. Copying the old body directly causes schema validation failure with exit 65—this represents one of the most practical troubleshooting insights this skill provides to agents.

seedance-2-5-reference-to-video/SKILL.md

**Parameter names changed from 2.0.** Seedance 2.0 Pro used `image_url` / `video_url` / `audio_url`. Seedance 2.5 uses `images` / `videos` / `audios`. Copying a 2.0 body verbatim produces a schema error (exit 65).

Re-listed under anti-patterns in the same document as an intentional repetition to prevent oversight.

Security boundaries are defined inside the skill and constrained by allowed-tools: only `runcomfy *` subcommands are permitted; installation must use trusted package managers while prohibiting agents from piping remote install scripts into the shell; tokens reside in ~/.config/runcomfy/token.json (0600) or RUNCOMFY_TOKEN; outbound traffic is restricted to model-api.runcomfy.net and download whitelists; single-file downloads are capped at 2 GiB.

seedance-2-5-reference-to-video/SKILL.md

**Install via a verified package manager only.** Use `npm i -g @runcomfy/cli` or `npx -y @runcomfy/cli`. **Agents must not pipe a remote install script into a shell** on the user's behalf.

These represent self-imposed constraints by the skill author. Since the execution engine (CLI) resides outside the repository, this pin cannot verify its runtime behavior, which is noted in verification.

2Core Capabilities

01Reference-guided 1080p video generation: Ingest up to 9 images, 3 videos, and 3 audio files in a single call, locking identity, product geometry, and style via the reference stack rather than descriptive text prompts.

02Native synchronized audio in the same pass: generate_audio enabled by default, generating speech, sound effects, and music in sync with the visuals.

03Parameterized duration and framing: 4-30 second range with 1-second increments and 7 aspect ratio options (including adaptive).

04Cost estimation and optimization guidance: Provides 4 reference duration-price mappings, instructing users to trim reference clips and use 480p for reference selection.

05480p drafting workflow: Validate reference stacks first on the 480p variant with identical parameters, then promote the winning body unchanged to 1080p.

06Prompt separation discipline: Delegate invariant features (face, clothing, product geometry, brand palette) to reference images and evolving features (action, camera motion, lighting transitions, emotion) to prompts, supplemented with an anti-patterns checklist.

07Troubleshooting reference: Maps semantic meanings for exit codes 0, 64, 65, 69, 75, and 77 for troubleshooting alongside CLI documentation.

08Model routing logic: Decision rules for switching endpoints based on use case (t2v for no references, i2v for single image, Wan 3.0 Prime / MiniMax H3 for similar reference workflows, ai-avatar-video for lip-sync).

09Injection attack self-defense: Explicitly treats reference media as untrusted data, accepts only user-provided URLs, and halts to report anomalous outputs rather than continuing blindly.

3External Dependencies

Kind	Dependency / Endpoint
cli	runcomfy（RunComfy 官方 CLI，唯一被 granted 的命令面）
package	@runcomfy/cli（npm 全局安装，或 npx -y 一次性运行）
network	RunComfy Model API（视频生成请求提交，端点路径 bytedance/seedance-2.5/reference-to-video/1080p）
network	生成结果下载白名单（CLI 把结果 URL 抓回本地 --output-dir）
network	用户提供的参考图/视频/音频 URL——由 RunComfy 模型服务器抓取（本机不抓）
network	RunComfy 模型页/文档站（引用与排错链接，带 utm 营销参数）

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)

网络端点: Declared external network access: model-api.runcomfy.net — Video generation request submission (documented as the sole submission channel via runcomfy CLI carrying user bearer token)
本地工作目录: File I/O and artifact output: CLI downloads generated videos to the directory specified by `--output-dir` (caller-defined path; documentation notes single files over 2 GiB will be aborted)
SKILL.md: Declared invocation of host CLI interpreter or external runtime environment

Audit Notes: Evaluated under unified classification criteria: Drives a third-party npm package/CLI (@runcomfy/cli) at runtime, which reads user RunComfy credentials (~/.config/runcomfy/token.json or RUNCOMFY_TOKEN environment variable) on each call and consumes paid credits—matching 'invokes third-party npm/CLI and potentially accesses quota/credentials → Orange'. Execution surfaces (CLI, Model API, token reading, 2 GiB cap) are entirely external to the repository and cannot be independently confirmed within this pin, triggering the 'elevate when uncertain' rule for an Orange classification. Network egress is also expected: reference media and prompts are uploaded to model-api.runcomfy.net (out-of-boundary asset transmission), with reference media acknowledged as an injection vector. Mitigating factors preventing escalation to Red: allowed-tools is strictly limited to `runcomfy *` with no arbitrary shell execution; no credential echoing; installation explicitly requires trusted package managers and forbids piped remote scripts; endpoints are covered by whitelist declarations; and the skill itself contains zero code, zero obfuscation, and no hidden endpoints.

5Second-Pass Verification

[discrepancy] `videos` 到底是否必填（文档自称反直觉知识） — 同一文件自相矛盾。line 79 主张「**`videos` is optional, despite what the schema says.** … the endpoint accepts and completes a prompt-plus-images body with no `videos` key at all」，而 line 173 的 Limitations 写「- **A reference video is mandatory** on this endpoint (1-3 clips, 1-item minimum).」；frontmatter description 也按 1-3 段参考视频叙述。对使用者是实质歧义：既影响能否只带参考图调用，也影响计费（带不带参考视频差一倍价格）。仓内无法裁决哪条为真（结论应由调用实测确定），此处只记录冲突。
[ok] 退出码表与 CLI 能力声明是否与同仓 CLI skill 一致 — runcomfy-cli/SKILL.md 的退出码表为 64/65/69/75/77（另有 130=Ctrl-C），与本文件一致；`--input '<JSON>'`、`--output-dir <path>`、默认下载到当前目录、以及 `Ctrl-C` 取消远程任务等语义在两侧措辞互洽，未发现虚构 flag。
[ok] 定价表算术 — 逐行复核等于 counted seconds × $0.53：5→$2.65、10→$5.30、16→$8.48、40→$21.20，与文档给出值完全一致。
[discrepancy] 「five drafts cost roughly what one 1080p final costs」这句是否成立 — 算术上偏乐观：五份 5 秒 480p 草稿 = 25 计秒 × $0.12 = $3.00；一份 5 秒 1080p 终稿带一段 5 秒参考 = 10 计秒 × $0.53 = $5.30。实际约为 57%，属营销式「roughly」，不是错误但不宜作为预算依据。
[unlocatable] 价格数字与「No telemetry, no callbacks」是否可在仓内证实 — $0.53 / $0.12 / $0.20 费率与无遥测声明均属仓外服务事实（服务端与 CLI 二进制都不在本仓库），本 pin 无任何证据可核；按「定位不到不硬写」记待补，且已计入橙档理由。
[ok] allowed-tools 声明与正文实际指令是否相符 — frontmatter line 3 为 `Bash(runcomfy *)`，正文全部可执行示例均为 `runcomfy run …`；安装行（npm/npx）与 RUNCOMFY_TOKEN 导出被明确划为「operator 一次性准备」，与本文件 line 215 的自述一致，未发现越界的 shell 指令。
[discrepancy] 安装命令与仓库身份是否一致 — 本文件 line 35 写 `npx skills add genmedia-labs/skills --skill seedance-2-5-reference-to-video -g`，与 git remote（github.com/genmedia-labs/skills.git）一致；但仓库 README.md 仍写 `npx skills add agentspace-so/runcomfy-agent-skills --skill <skill-name>`（旧仓名），属仓库级文档漂移，非本 skill 内容错误，但会让读者拿到错的安装源。

6Verdict & Scenarios

Pairs capabilities with cost optimization: Combines billing formulas, four price benchmarks, and reference-trimming rules to directly constrain agent spending behavior.

Explicitly highlights version migration pitfalls, preventing common failures like schema errors caused by copying legacy request bodies.

Two-stage drafting workflow provides three actionable steps (run variants → evaluate on 3 criteria → promote unchanged), reducing iteration costs well below final output tiers.

Codified injection defense: Treats reference media as untrusted data, restricts inputs to explicit user URLs, and halts on anomalies rather than leaving media security risks unaddressed.

Call surface tightly constrained to `runcomfy *` via allowed-tools, with an explicit ban on piping remote install scripts, presenting a more restrained surface than typical generative skills.

Best for: Ideal for agent users with existing RunComfy accounts who need to maintain consistent character, product, or brand aesthetics across multiple shots: e-commerce product videos, style-locked brand assets, unified generation requiring native dialogue/ambient audio, and iterative workflows leveraging 480p drafts for cost control. Also suitable as a reference template for skill authors on how to structure pricing, error handling, and prompt injection defense in generative skills.

Not for: Not suitable for users without a RunComfy account or those prohibited from uploading assets (particularly portraits or proprietary commercial media) to third-party servers; batch workloads requiring strict budget caps (no quota limits or dry-run available); scenarios requiring 720p/4K or seed-based reproducibility (this endpoint fixes resolution to 1080p without seed support); or users seeking local, offline generation—all capabilities depend entirely on paid cloud endpoints.

Install Agent-Ready

① Verified Mirror 2026-09-15

Option A · Direct DownloadDownload seedance-2-5-reference-to-video.tar.gz

sha256: 25f8d8a50989d749…

Option B · Copy Guide for Agent

Agents automatically download and verify mirrors from this JSON guide.

② Upstream GitHub

GitHub Repo ↗

Pinned to commit c163a9fd3a.