Vidu参考生视频

Vidu参考生视频 - "Vidu reference video" - is the middle child of the pack's three Vidu nodes, and it's the one people usually mean when they say "I want video from my existing images." The images input (a batch of IMAGEs) acts as the visual anchor: the model generates a video that follows the look, character, or product in those references. It's the mode to reach for when you have a specific subject - a product, a character, a face - and you want it moving in a scene, without the rigidity of first/last-frame animation.

The mechanism is consistent with the other two: images are base64'd and POSTed to Mojie's endpoint under the vidu_video model (the viduq2 backend), the result is downloaded to a temp file, and it lands in the graph as a VIDEO type via VideoFromFile. Nothing runs locally, which is the pack's whole bargain - you're renting Vidu's motion understanding instead of installing a local video model that would laugh at a low-end GPU.

The inputs that matter

• images - required, and it's the star. A batch of reference images (the pack's MultiImageUpload node feeds this beautifully). The more consistent the references, the more consistent the video.
• prompt - what happens in the video. Describing motion helps; describing the subject is redundant since the images carry it.
• aspect_ratio - 16:9, 9:16, or 1:1.
• duration - 1 to 8 seconds, default 5.
• resolution - 720p or 1080p (the cost lever).
• movement_amplitude - auto / small / medium / large; how much motion you want.
• seed - here's the quirk: the source hardcodes the payload seed to 0, so changing this input won't change your output. The field exists and does nothing on this node. Don't waste time wiggling it for variety - that's a known papercut in the code, not a bug you can fix from the canvas.

Output is video.

Installing it

The pack install, third time now, because it's one repo:

cd ComfyUI/custom_nodes
git clone https://github.com/MoJIeAIGC/comfyui-MJAPI-party.git

or ComfyUI Manager search "mojieapi_party". Key from mojieaigc.com into config.ini, restart. No models to download; opencv-python is the only heavy dependency.

Honest caveats

• Seed does nothing here. Covered above - the dead input is a real gotcha for anyone who expects Vidu to behave like a local sampler.
• Cost. Video calls are the pack's most expensive category, and the KB's API-nodes essay documents the community's sticker shock at hosted video pricing. 1080p and long durations multiply the burn; preview cheap at 720p/5s before committing to a final render.
• Reference drift. If your subject's identity wobbles between shots, tighten the references - consistent framing and lighting in the input images beats any prompt. Your images also leave the machine, so keep that in mind for unreleased product.