seedance

Generate videos with the doubao-seedance-2.0 model (text-to-video, image-to-video, first/last-frame-to-video). Use when the user asks to "generate a video", mentions "seedance", or wants AI video generation from text and/or images. Wraps the asynchronous Volcengine Ark / OpenAI-style video API behind one synchronous call.

seedance — doubao-seedance-2.0 video generation

This skill generates a video with the doubao-seedance-2.0 model by running scripts/seedance.py (paths below are relative to this skill's directory, i.e. the folder containing this SKILL.md). The script is synchronous: it creates a task, polls until completion, and prints the result (including the video URL).

When to use

Trigger this skill when the user's request mentions "seedance" or asks to "generate a video" / produce AI video from text, an image, or first/last frames. Also trigger on explicit /seedance.

Do not trigger for image-only generation, audio, or general chat.

Prerequisites (check before running)

• Python 3.11+. The entrypoint wrapper (scripts/seedance) auto-selects python3.11/python3.12 when available (some proxies' TLS cert chain is rejected by Python 3.14; python3 is fine for the raw Ark API), so you do not pick the interpreter by hand.
• The same wrapper creates an isolated virtualenv at <skill_dir>/.venv on first run and installs requirements.txt there automatically. The only Python dependency is alibabacloud_oss_v2 — pulled in only when the user supplies local image files (uploaded to Alibaba OSS as a signed URL), but harmless to install regardless. No manual pip install is needed.
• ARK_API_KEY is required. It may be a real env var or stored in a .env file (in the skill dir or the user's cwd). If absent, ask the user for it and offer to write it to <skill_dir>/.env (never print the value back).
• Optional env overrides: ARK_API_TYPE (ark/openai-video/openai), ARK_MODEL (model id), ARK_ENDPOINT. All are lower priority than their CLI flags (--api-type, -m, --endpoint). For local images also: OSS_ACCESS_KEY_ID, OSS_ACCESS_KEY_SECRET (required), and optionally OSS_ENDPOINT, OSS_BUCKET (default jiangsier), OSS_KEY_PREFIX (default dev/).

Extract options from the user's query

Map the user's natural-language request to these CLI options:

User intent	CLI option
The video prompt / description	-t / --text (required)
A local reference image file	-i / --image
An online reference image URL	--image-url
A local first-frame image	-f / --first-frame
An online first-frame URL	--first-frame-url
A local last-frame image	-l / --last-frame
An online last-frame URL	--last-frame-url
"5 seconds", "10s", duration	--duration (default 5)
"16:9", "vertical/9:16", "square/1:1", "follow the image"	--ratio (default 16:9; use adaptive when an image is provided and the user wants the output to match it)
"720p", "1080p", "4k"	--resolution (default 720p)
A specific model id	-m / --model (or ARK_MODEL env)
Which backend protocol	--api-type (default ark; or ARK_API_TYPE env). ark = raw Volcengine Ark (audio). openai-video = openai-video endpoint, doubao-seedance-2.0-mini (silent). openai = openai chat-mirror entry, doubao-seedance-2.0 full (silent).
Reproducible result	--seed (ark/openai only; openai-video ignores it)
Save the mp4 locally	--save <path>

Rules (enforced by the script; surface them to the user when relevant):

• --image and --image-url are mutually exclusive (same for first/last frame).
• A reference image (--image/--image-url) cannot be combined with explicit first/last frames.
• --last-frame requires --first-frame.
• --ratio adaptive requires an image input.

Ask proactively, then confirm

1. Prompt: if the user's request does not contain a usable video prompt, ask for one (it is required).
2. API key: if ARK_API_KEY is not in the environment or a .env, ask the user for it (offer to store it in <skill_dir>/.env).
3. Ambiguous image role: if the user gives an image without saying whether it is a reference, first frame, or last frame, ask.
4. Confirm before running — generation costs quota and runs for minutes. Summarize the resolved options and use AskUserQuestion (or just ask) to get explicit confirmation, including: prompt, image inputs, duration, ratio, resolution, model, api-type. Proceed only after the user confirms.

Invocation

Resolve <skill_dir> (this directory). Load secrets from <skill_dir>/.env if present via --env-file. Run the skill through the scripts/seedance wrapper — it provisions an isolated venv (with requirements.txt installed) and forwards every argument to scripts/seedance.py, picking python3.11/python3.12 for the proxy-TLS-safe interpreter automatically. The same wrapper serves all three --api-type values; no interpreter switch is needed between them.

bash

# Ark (default) — also covers openai-video / openai; just pass --api-type
<skill_dir>/scripts/seedance --env-file <skill_dir>/.env \
  -t "<prompt>" [-i ./local.png | --image-url <url>] [-f ./first.png -l ./last.png] \
  [--duration 5] [--ratio 16:9] [--resolution 720p] [-m <model>] [--save out.mp4]

<skill_dir>/scripts/seedance --api-type openai-video --env-file <skill_dir>/.env -t "<prompt>" ...
<skill_dir>/scripts/seedance --api-type openai      --env-file <skill_dir>/.env -t "<prompt>" ...

The script prints the full task result as JSON on stdout and the video_url on stderr. On success, report the video URL to the user; if --save was given, confirm the saved path.

To run the test suite (offline, no key needed): <skill_dir>/scripts/seedance test_seedance.py. The wrapper reuses the same venv for any script in scripts/ — pass its filename as the first argument.

Examples

• "Generate a 5s video of a daisy field under a blue sky, camera pushing in" → scripts/seedance -t "a daisy field under a blue sky, camera pushing in" --duration 5
• "Make a video from this image ./fox.png — the camera slowly pulls out, hair blowing in the wind" → scripts/seedance -t "camera slowly pulls out, hair blowing in the wind" -i ./fox.png --ratio adaptive
• "Animate between these two frames: ./first.jpeg and ./last.jpeg, 360° orbit" → scripts/seedance -t "360 degree orbit" -f ./first.jpeg -l ./last.jpeg --ratio adaptive
• "Generate a vertical 1080p 10s clip of waves at sunset" → scripts/seedance -t "waves at sunset" --ratio 9:16 --resolution 1080p --duration 10

Notes

• Fixed per-spec params (ark + openai): generate_audio=true, watermark=false (not exposed). The openai-video body omits these. Both openai-video and openai yield silent video on this backend — only ark produces audio.
• Three API protocols (--api-type, or ARK_API_TYPE env): ark (default; Volcengine Ark, POST /contents/generations/tasks, audio), openai-video (POST /video/generations, OpenAI-video body seconds/size/input_reference, model doubao-seedance-2.0-mini, silent), and openai (POST /video/generations, Ark-shaped body, model doubao-seedance-2.0 full, silent — the proxy's chat-mirror entry; /v1/chat/completions returns unsupported_model_endpoint here). The script normalizes all three into {id, status, video_url, error, raw}.
• Default model id: doubao-seedance-2-0-260128 (ark) / doubao-seedance-2.0-mini (openai-video) / doubao-seedance-2.0 (openai). Override with -m or ARK_MODEL.
• reference.md (in this skill's directory) is the upstream Volcengine Ark video-generation doc; consult it for advanced/parameter details.
• Local images are uploaded to Alibaba OSS with a 10-minute signed URL, then that URL is used as the model input.