Vidu Reference To Video Generation

Vidu Reference to Video is the multi-image video generator: feed it up to seven images plus a prompt, and Vidu (the Chinese lab's video model, API-only like most of the peer tier) produces a clip where the same subjects stay consistent across the shots. Where a single first-frame node locks in one composition, this one lets you hand over several angles of a character or product and have the model keep that identity while the camera and action move. It's a built-in partner node - cloud render through Comfy's API, billed per call, no local model.

This is the node to reach for when character or product consistency is the whole point: a character sheet of a few frames, a product from multiple angles, and the output video respects all of them instead of inventing a look. The cost is real (roughly the pricey end of the i2v tier), so it's for when the reference fidelity justifies the spend.

The inputs

• images - your reference images, max 7. More than that and the node rejects the run. Each image is validated: at least 128×128px and an aspect ratio between 1:4 and 4:1, so don't feed it panoramic or tiny crops.
• prompt - the textual description of the video you want.
• model - the Vidu model (viduq1).
• aspect_ratio - 16:9, 9:16, or 1:1.
• movement_amplitude - how much motion the model should put into the frame: auto, small, medium, large. Small = calm, controlled shots; large = energetic camera and action. This is the knob people forget and then wonder why everything's static.
• duration - fixed at 5 seconds (the schema pins it; don't go looking for a longer option on this node).
• seed (optional) - 0 for random, otherwise a fixed seed for the attempt.
• resolution - 1080p.

What comes out

A single VIDEO output at 1080p.

Gotchas

• Keep an eye on billing. Vidu's API through Comfy has a real track record of credit surprises - there are community threads from people watching their balance evaporate after a couple of Vidu calls, with the template-advertised price not matching the deduction. Check the price badge on the node, and spot-check your credit usage after the first render before you batch.
• Consistent references, consistent results. The images are the consistency mechanism; mismatched lighting or framing across them and the model blurs the differences instead of locking identity.
• The 7-image and ratio limits are enforced. Overshoot and it fails (before billing, at least).
• Seed ≠ reproducibility, as always with this family - it's a per-attempt input, not a guarantee you can replay.

The workflow that clicks: a few frames of the same subject, a prompt that says what happens, movement_amplitude set on purpose - and a video that finally keeps your character looking like your character.