Category: provider
Model Studio Wan Video
Validation
mkdir -p output/aliyun-wan-video
python -m py_compile skills/ai/video/aliyun-wan-video/scripts/generate_video.py && echo "py_compile_ok" > output/aliyun-wan-video/validate.txt
Pass criteria: command exits 0 and output/aliyun-wan-video/validate.txt is generated.
Output And Evidence
- Save task IDs, polling responses, and final video URLs to
output/aliyun-wan-video/.
- Keep one end-to-end run log for troubleshooting.
Provide consistent video generation behavior for the video-agent pipeline by standardizing video.generate inputs/outputs and using DashScope SDK (Python) with the exact model name.
Critical model names
Use one of these exact model strings:
wan2.6-t2vwan2.6-t2v-uswan2.2-t2v-pluswan2.2-t2v-flashwan2.6-i2v-flashwan2.6-i2vwan2.6-i2v-uswanx2.1-t2v-turbo
Prerequisites
- Install SDK (recommended in a venv to avoid PEP 668 limits):
python3 -m venv .venv
. .venv/bin/activate
python -m pip install dashscope
- Set
DASHSCOPE_API_KEY in your environment, or add dashscope_api_key to ~/.alibabacloud/credentials (env takes precedence).
Normalized interface (video.generate)
Request
prompt (string, required)negative_prompt (string, optional)duration (number, required) secondsfps (number, required)size (string, required) e.g. 1280*720seed (int, optional)reference_image (string | bytes, optional for t2v, required for i2v family models)motion_strength (number, optional)
Response
video_url (string)duration (number)fps (number)seed (int)
Quick start (Python + DashScope SDK)
Video generation is usually asynchronous. Expect a task ID and poll until completion.
Note: Wan i2v models require an input image; pure t2v models such as wan2.6-t2v can omit reference_image.
import os
from dashscope import VideoSynthesis
# Prefer env var for auth: export DASHSCOPE_API_KEY=...
# Or use ~/.alibabacloud/credentials with dashscope_api_key under [default].
def generate_video(req: dict) -> dict:
payload = {
"model": req.get("model", "wan2.6-i2v-flash"),
"prompt": req["prompt"],
"negative_prompt": req.get("negative_prompt"),
"duration": req.get("duration", 4),
"fps": req.get("fps", 24),
"size": req.get("size", "1280*720"),
"seed": req.get("seed"),
"motion_strength": req.get("motion_strength"),
"api_key": os.getenv("DASHSCOPE_API_KEY"),
}
if req.get("reference_image"):
# DashScope expects img_url for i2v models; local files are auto-uploaded.
payload["img_url"] = req["reference_image"]
response = VideoSynthesis.call(**payload)
# Some SDK versions require polling for the final result.
# If a task_id is returned, poll until status is SUCCEEDED.
result = response.output.get("results", [None])[0]
return {
"video_url": None if not result else result.get("url"),
"duration": response.output.get("duration"),
"fps": response.output.get("fps"),
"seed": response.output.get("seed"),
}
Async handling (polling)
import os
from dashscope import VideoSynthesis
task = VideoSynthesis.async_call(
model=req.get("model", "wan2.6-i2v-flash"),
prompt=req["prompt"],
img_url=req["reference_image"],
duration=req.get("duration", 4),
fps=req.get("fps", 24),
size=req.get("size", "1280*720"),
api_key=os.getenv("DASHSCOPE_API_KEY"),
)
final = VideoSynthesis.wait(task)
video_url = final.output.get("video_url")
Operational guidance
- Video generation can take minutes; expose progress and allow cancel/retry.
- Cache by
(prompt, negative_prompt, duration, fps, size, seed, reference_image hash, motion_strength).
- Store video assets in object storage and persist only URLs in metadata.
reference_image can be a URL or local path; the SDK auto-uploads local files.- If you get
Field required: input.img_url, the reference image is missing or not mapped.
wan2.6-t2v and wan2.6-t2v-us add multi-shot narrative support and optional audio input according to the official docs.
Size notes
- Use
WxH format (e.g. 1280*720).
- Prefer common sizes; unsupported sizes can return 400.
Output location
- Default output:
output/aliyun-wan-video/videos/
- Override base dir with
OUTPUT_DIR.
Anti-patterns
- Do not invent model names or aliases; use official Wan i2v model IDs only.
- Do not block the UI without progress updates.
- Do not retry blindly on 4xx; handle validation failures explicitly.
Workflow
- Confirm user intent, region, identifiers, and whether the operation is read-only or mutating.
- Run one minimal read-only query first to verify connectivity and permissions.
- Execute the target operation with explicit parameters and bounded scope.
- Verify results and save output/evidence files.
References
