Video Generate Skill

This skill generates videos using Doubao Seedance 1.0/1.5 models.

Trigger Conditions

1. User wants to generate videos from text descriptions
2. User wants to create videos based on images (first/last frame)
3. User wants to create videos with reference materials (images, videos, audio)
4. User asks for video generation capabilities

Usage

Environment Variables

Before using this skill, ensure the following environment variables are set:

• ARK_API_KEY or MODEL_VIDEO_API_KEY or MODEL_AGENT_API_KEY: API key for the video generation service
• MODEL_VIDEO_API_BASE: API base URL (optional, has default)
• MODEL_VIDEO_NAME: Model name (optional, has default)

Function Signature

async def video_generate(
    params: list,
    batch_size: int = 10,
    max_wait_seconds: int = 1200,
    model_name: str = None,
) -> Dict:

Parameters

params (list[dict])

A list of video generation requests. Each item is a dict with the following fields:

Required per item:

• video_name (str): Name/identifier of the output video file
• prompt (str): Text describing the video to generate. Supports Chinese and English.

Optional per item - Input Materials:

• first_frame (str): URL for the first frame image
• last_frame (str): URL for the last frame image
• reference_images (list[str]): 1-4 reference image URLs for style/content guidance
• reference_videos (list[str]): 0-3 reference video URLs (mp4/mov, 2-15s each, total ≤15s)
• reference_audios (list[str]): 0-3 reference audio URLs (mp3/wav, 2-15s each, total ≤15s)

Optional per item - Video Output Parameters:

• ratio (str): Aspect ratio. Options: "16:9" (default), "9:16", "4:3", "3:4", "1:1", "2:1", "21:9", "adaptive"
• duration (int): Video length in seconds. Range: 2-12s depending on model
• resolution (str): Video resolution. Options: "480p", "720p", "1080p"
• frames (int): Total frame count. Must be in [29, 289] and follow format 25 + 4n
• camera_fixed (bool): Lock camera movement. Default: false
• seed (int): Random seed for reproducibility. Range: [-1, 2^32-1]
• watermark (bool): Whether to add watermark. Default: false
• generate_audio (bool): Whether to generate audio. Only Seedance 1.5 supports this
• tools (list[dict]): Tool configuration, e.g., [{"type": "web_search"}]

Input Modes

1. Text-to-Video: Only provide prompt, no images/videos
2. First Frame Guidance: Provide first_frame for starting image
3. First + Last Frame Guidance: Provide both for transition video
4. Reference Images: Provide reference_images for style/content guidance
5. Multimodal Reference: Combine reference_images, reference_videos, reference_audios

Return Value

Script Return Info

The video_generate.py script will return these info:

{
    "status": "success" | "partial_success" | "error",
    "success_list": [{"video_name": "video_url"}],
    "error_list": ["video_name"],
    "error_details": [{"video_name": "...", "error": {...}}],
    "pending_list": [{"video_name": "...", "task_id": "cgt-xxx", ...}]
}

Based on the script return info, the final response returned to the user consists of a description of the video generation task and the video URL(s). You may download the video from the URL, but the video URL should still be provided to the user for viewing and downloading.

Note: the URL is the 'url' in the success_list of script return info. The URL must return in two ways:

Final Return Info

For OpenClaw

1. First, save/download the generated video to an allowed directory:

   • Recommended: ~/.openclaw/workspace/

2. Use the message tool to send the video with these parameters:

   {
     "action": "send",
     "message": "Optional text description",
     "media": "/root/.openclaw/workspace/generated-video.mp4"
   }
   

   • The tool will automatically: read the local file, upload to the channel server as a web URL, and send as a video message

3. Verify success: Check that the tool returns {"ok": true} to confirm the video was sent successfully

4. Normal Text not message tool: After generation, show list of videos with Markdown format, for example：

   <video src="https://example.com/video1.mp4" width="640" controls>video-1</video>
   

   Very important: The video URL must be an original online resource link starting with https://, not a local video address you have downloaded.

For Normal

You should return three types of information:

1. File format, return the video file (if you have some other methods to send the video file) and the local path of the video, for example: local_path: /root/.openclaw/workspace/skills/video-generate/xxx.mp4
2. After generation, show list of videos with Markdown format, for example:

<video src="https://example.com/video1.mp4" width="640" controls>video-1</video>
<video src="https://example.com/video2.mp4" width="640" controls>video-2</video>

Code Implementation

See scripts/video_generate.py for the full implementation.

Example Usage

# Text-to-Video
python scripts/video_generate.py -p "小猫骑着滑板穿过公园" -n cat_park -r 16:9 -d 5 --resolution 720p

# First Frame Guidance
python scripts/video_generate.py -p "小猫跳起来" -n cat_jump -f "https://example.com/cat.png" -r adaptive -d 5

# First + Last Frame Guidance
python scripts/video_generate.py -p "平滑过渡动画" -n transition \
    -f "https://example.com/start.png" \
    -l "https://example.com/end.png" \
    -d 6

# Reference Images (style/content guidance)
python scripts/video_generate.py -p "[图1]戴着眼镜的男生和[图2]柯基小狗坐在草坪上" -n styled \
    --ref-images "https://example.com/boy.png" "https://example.com/dog.png" \
    -r 16:9 -d 5

# Multimodal Reference (video + audio)
python scripts/video_generate.py -p "将视频中的人物换成[图1]中的男孩" -n multimodal \
    --ref-images "https://example.com/boy.png" \
    --ref-videos "https://example.com/source.mp4" \
    --ref-audios "https://example.com/voice.wav" \
    -d 5

# With Audio Generation (Seedance 1.5 only)
python scripts/video_generate.py -p "女孩抱着狐狸，可以听到风声和树叶沙沙声" -n with_audio \
    -f "https://example.com/girl_fox.png" \
    --generate-audio \
    -m doubao-seedance-1-5-pro-251215 \
    -d 6 --resolution 1080p

# Query task status
python scripts/video_generate.py -q "cgt-20260222165751-wsnw8"

# Use specific model
python scripts/video_generate.py -p "A futuristic city" -m doubao-seedance-2-5-260628

# No watermark
python scripts/video_generate.py -p "A beautiful landscape" --no-watermark

Command Line Options

Option	Short	Description
--prompt	-p	Text description of the video (required)
--name	-n	Video name identifier (default: video)
--model	-m	Model name (default: doubao-seedance-2-5-260628)
--ratio	-r	Aspect ratio (default: 16:9)
--duration	-d	Video duration in seconds (2-12)
--resolution		Video resolution: 480p, 720p, 1080p
--first-frame	-f	First frame image URL
--last-frame	-l	Last frame image URL
--ref-images		Reference image URLs (space-separated, 1-4 images)
--ref-videos		Reference video URLs (space-separated, 0-3 videos)
--ref-audios		Reference audio URLs (space-separated, 0-3 audios)
--generate-audio		Generate audio (Seedance 1.5 only)
--seed		Random seed for reproducibility
--no-watermark		Disable watermark
--timeout	-t	Max wait time in seconds (default: 1200)
--query-task	-q	Query task status by task_id

Model Fallback

If you encounter a model-related error (like ModelNotOpen), you can downgrade to these models:

• doubao-seedance-1-5-pro-251215
• doubao-seedance-1-0-pro-250528

Error Handling

• IF the script raises the error "PermissionError: ARK_API_KEY or MODEL_VIDEO_API_KEY or MODEL_AGENT_API_KEY not found in environment variables", inform the user that they need to provide the ARK_API_KEY or MODEL_VIDEO_API_KEY or MODEL_AGENT_API_KEY environment variable. Write it to the environment variable file in the workspace. If the file already exists, append it to the end. Ensure the environment variable format is correct, make the environment variable effective, and retry the video generation task that just failed.

Notes

• Keep prompt concise (recommended ≤ 500 characters)
• For first/last frame, ensure aspect ratios match your chosen ratio
• Reference images: 1-4 images, formats: jpeg/png/webp/bmp/tiff/gif
• Reference videos: 0-3 videos, formats: mp4/mov, total duration ≤ 15s
• Reference audios: 0-3 audios, formats: mp3/wav, total duration ≤ 15s
• Multimodal requires at least one image or video (audio-only not supported)
• Audio generation is only supported by Seedance 1.5 pro
• If polling times out, use --query-task with the returned task_id