Volcengine Video Studio

Production-ready Volcengine/ARK video generation for prompt-to-video, image-to-video, and draft-video refinement. Use when users want a hands-on skill that c...

Security Scan

The skill appears to do what it says (submit/poll/download Volcengine/ARK video jobs); environment and local-file use are expected for this purpose, though the package metadata omitted declared env requirements.

Purpose & Capability

Name/description match the included script and instructions: the skill submits video generation tasks to a Volcengine/ARK API, polls status, inspects payloads, and downloads outputs. Requiring an API key and endpoint is proportional to the stated purpose.

Instruction Scope

SKILL.md and the script explicitly read local media files (encode as data: URLs), assemble request payloads, post to the declared endpoints, poll for task status, and download resulting files to a local directory. These actions are expected for a real video-generation runner but do mean local files and prompts will be transmitted to the remote service and downloaded into ~/Desktop/volcengine-videos by default. The skill does not attempt to read unrelated system files or other credentials.

Install Mechanism

There is no install spec; this is instruction + bundled script only. Nothing is downloaded from external URLs at install time and no archives are extracted.

Credentials

The script legitimately needs an API key and endpoint (VOLCENGINE_API_KEY/ARK_API_KEY and VOLCENGINE_VIDEO_ENDPOINT / ARK_BASE_URL). However, the registry metadata lists no required env vars while SKILL.md and the script do — this mismatch is noteworthy. The script also falls back to OPENAI_API_KEY / OPENAI_BASE_URL as alternative env names, which is unusual but seems intended as convenience/fallback; verify you want that behavior before supplying an OpenAI key in a shared environment.

Persistence & Privilege

always is false; the skill does not request persistent presence or modify other skills. It runs as a user-invoked script and can be called autonomously by the agent (normal default) but it does not ask for elevated system privileges.

Assessment

This skill is internally consistent with a Volcengine/ARK video runner, but note the following before installing: (1) You must provide a Volcengine/ARK API key and endpoint (the README lists these), even though the registry metadata omitted them. (2) The script will read local image/video files, base64-encode them, and send them to the remote API — only use with media you are comfortable uploading. (3) The code will also download generated outputs to ~/Desktop/volcengine-videos by default; set --download-dir if you prefer a different location. (4) The script will fall back to OPENAI_API_KEY/OPENAI_BASE_URL environment names if Volcengine vars are not set — avoid having unrelated keys in your environment if you don't want them reused. If you require higher assurance, request the full untruncated script for review and confirm the exact network endpoints the tool will call.

Use this skill to actually submit, poll, inspect, and download Volcengine / ARK video generation jobs instead of only drafting prompts.

中文说明

这是一个面向 火山引擎 / ARK 兼容视频生成接口 的生产可用型 skill，重点不是只写视频提示词，而是把整条生成链路真正跑通，从提交任务、轮询状态，到拿回结果、下载成片，都能一把做完。

它特别适合下面这些对外场景：

• 文生视频
• 图生视频
• 草稿视频二次优化
• 提交生成任务并持续轮询状态
• 检查任务返回内容与原始 payload
• 将生成好的视频自动下载到本地

如果你想要的是一个能真实发起任务、能追踪任务进度、能拿到最终视频文件、适合实战交付的火山视频 skill，这个版本就是为这种需求准备的。

Treat the prompt and optional reference media as runtime input for each task. Do not reuse documentation example prompts or example media in a real generation request unless the user explicitly asks for that exact example.

What this skill is good at

• text-to-video from a task-specific prompt
• image-to-video using a local file, data URL, or remote image URL
• draft-video refinement with a guiding prompt
• polling existing tasks and downloading finished outputs
• switching between a faster default model and a stronger alternate model

Default path

Run the bundled script with the actual prompt for the current task:

python3 scripts/generate_video.py "当前任务的实际视频提示词"

To switch to the alternate model for one run:

VOLCENGINE_VIDEO_MODEL=doubao-seedance-1-5-pro-251215 \
python3 scripts/generate_video.py "当前任务的实际视频提示词"

By default the script:

• submits the task
• polls until the task finishes
• extracts returned video URLs from the task payload
• downloads generated files into ~/Desktop/volcengine-videos/<timestamp>-<slug>/

Required config

The script reads config from env vars:

• VOLCENGINE_API_KEY or ARK_API_KEY
• VOLCENGINE_VIDEO_MODEL (recommended)
• VOLCENGINE_VIDEO_ENDPOINT or VOLCENGINE_ENDPOINT or ARK_BASE_URL

Recommended video models:

• doubao-seedance-1-0-pro-fast-251015 — default, faster iteration
• doubao-seedance-1-5-pro-251215 — alternate higher-tier option

Default behavior:

• if VOLCENGINE_VIDEO_MODEL is unset, the script defaults to doubao-seedance-1-0-pro-fast-251015
• avoid relying on VOLCENGINE_MODEL for video runs when that env var is also used for image generation

Typical endpoint:

• https://ark.cn-beijing.volces.com/api/v3

The script calls:

• POST /contents/generations/tasks
• GET /contents/generations/tasks/{task_id}

Common workflows

1. Text to video

python3 scripts/generate_video.py "为当前需求编写的视频提示词" \
  --ratio 16:9 \
  --duration 5

2. Image to video

python3 scripts/generate_video.py "描述你希望图片如何动起来" \
  --image ~/Desktop/reference.png \
  --ratio 9:16 \
  --duration 5

3. Draft/sample video to final video

python3 scripts/generate_video.py "说明要保留什么、增强什么" \
  --video ~/Desktop/draft.mp4

4. Inspect an existing task only

python3 scripts/generate_video.py --task-id <task_id> --wait false

5. Submit only, do not wait

python3 scripts/generate_video.py "当前任务的实际视频提示词" \
  --ratio 16:9 \
  --wait false

6. Use raw content JSON when the target model needs a custom body

python3 scripts/generate_video.py --content-json '[
  {"type":"text","text":"当前任务的实际视频提示词"},
  {"type":"image_url","image_url":{"url":"https://example.com/reference.png"}}
]'

Local media support

For --image and --video:

• local file path → converted to data: URL automatically
• https://... URL → sent as-is
• data:... URL → sent as-is

This makes local reference media usable without manual upload steps.

Key options

• --ratio 16:9|9:16|1:1|adaptive
• --duration <seconds>
• --frames <count>
• --seed <int>
• --resolution <value>
• --camera-fixed true|false
• --watermark true|false
• --callback-url <https-url>
• --poll-interval <seconds>
• --timeout <seconds>
• --download-results true|false
• --download-dir <path>
• --print-request

Execution checklist

1. Confirm whether the user wants text-to-video, image-to-video, or draft-video refinement.
2. Treat the prompt as task-specific runtime input; never carry over example prompt text or example media into a real run unless explicitly requested.
3. Choose prompt-first mode by default; use --content-json only when the API shape must be customized.
4. Pass local reference media directly with --image or --video.
5. Prefer --duration for whole-second clips and --frames only when finer control is required.
6. Poll by default so the final answer includes actual output URLs or downloaded files.
7. Mention the saved file paths when downloads are enabled.
8. If the API returns an unexpected structure, surface the raw JSON instead of guessing.

Troubleshooting

• Missing key → set VOLCENGINE_API_KEY
• Missing or wrong model → set VOLCENGINE_VIDEO_MODEL
• If a video run accidentally picks an image model (for example doubao-seedream-4-5) → explicitly set VOLCENGINE_VIDEO_MODEL instead of reusing VOLCENGINE_MODEL
• Missing endpoint → set VOLCENGINE_VIDEO_ENDPOINT
• 401/403 → key invalid or missing permission
• 404 → endpoint wrong or region mismatch
• 400 → unsupported model/parameter combination
• Task remains queued too long → check quota, rate limit, or model availability
• No obvious video URL in response → inspect raw

References

• references/sources.md
• references/api-notes.md ing VOLCENGINE_MODEL
• Missing endpoint → set VOLCENGINE_VIDEO_ENDPOINT
• 401/403 → key invalid or missing permission
• 404 → endpoint wrong or region mismatch
• 400 → unsupported model/parameter combination
• Task remains queued too long → check quota, rate limit, or model availability
• No obvious video URL in response → inspect raw

References

• references/sources.md
• references/api-notes.md