Vidu Reference2Video

The hardest problem in AI video is keeping the same character from frame one to frame sixty - and doing it on purpose. Vidu Reference2Video is the node in this pack built for exactly that. You hand it up to seven reference images of your subject (from different angles, or in different outfits), give it a prompt, and it generates video where the subject stays recognizable. No training, no LoRA, no identity-pinning workflow - just reference images sent up with the request.

That's a big deal for the character-consistency problem the KB keeps circling: normally you'd spend real effort building identity-preservation pipelines locally. This node does the "same person, different action" shot in one call. It's the one I'd point someone at for a talking-head video, a product shot with a consistent prop, or a character doing something new across multiple clips.

The inputs that matter

• client - standard wiring from the Modelverse Client or Secret Client.
• model - viduq3-turbo (default, "fast generation") or viduq2 (older). For a hosted tool, turbo is the sensible daily driver.
• prompt - the action. Default is make it dance; replace it with what you actually want the subject to do. Max 2000 characters.
• duration - 1–10 seconds.
• aspect_ratio - 16:9, 9:16, 3:4, 4:3, or 1:1.
• resolution - 540p, 720p (default), 1080p.
• Reference images: image1 through image7 (wire IMAGE tensors), or image_urls - a multiline text field, one URL per line, for web-hosted references. You can mix both; the node collects everything, base64-encodes the tensors, and trims to the first 7 with a warning if you overshoot.
• seed and bgm (adds background music).

Outputs are url and task_id; url feeds Modelverse Preview Video.

How it works

The node bundles all your reference images into the task input with a vidu_type: reference2video flag, submits it, and polls every 5 seconds until the clip is done. The consistency comes from the model reading all references together - which is exactly why you should feed it varied shots (different angles, expressions, maybe different clothing) rather than seven near-identical frames. One reference image of a face from one angle gives the model much less to lock onto than three good views.

Install

Pack install: ComfyUI Manager → search "ComfyUl-UCloud" → Install → restart, or git clone https://github.com/ucloud/comfyui-plugin into custom_nodes. No model downloads, light deps, API key from the Modelverse console.

Honest caveats: reference-to-video is the fiddliest of the Vidu modes, so budget a few tries before the subject locks in - that's normal, not a broken node. And everything here is metered, so those retries cost. Start with a single strong reference, confirm the subject reads, then add more.