What Vidu is
Vidu is a video generation model with a particular emphasis on reference-based generation: you supply images of a subject, and the model uses them to keep that subject recognisable across generated shots.
- Text to video and image to video.
- Reference-to-video, using supplied images to condition the subject.
- An emphasis on multi-shot consistency, which is the hardest problem in AI video.
What it does well
| Use | Fit | Note |
|---|---|---|
| Multi-shot character work | Worth evaluating | Test properly before relying on it |
| Keeping a product recognisable | Worth evaluating | Accuracy still requires a gate |
| Short narrative sequences | Moderate to strong | Depends heavily on reference quality |
| Single atmospheric shots | Fine, but not distinctive | Other tools are comparable |
Test the claim, do not assume it: reference conditioning is the capability most likely to matter to you and the one most likely to be overstated. Run a five-shot consistency test before committing.
Preparing reference images
- Front-facing, neutral pose or orientation, even lighting.
- Plain background — background detail leaks into generated scenes.
- Sharp and high resolution; reference quality caps output quality.
- One canonical image per subject, used everywhere.
- Consistent colour and no heavy stylisation, which would propagate into output.
A consistency-driven workflow
-
Lock one canonical reference per subject
This is the single highest-leverage step.
-
Write the subject description once and paste it identically
Paraphrasing between shots reintroduces drift.
-
Generate beat by beat
Four to six seconds per beat.
-
Chain approved frames
Use the best frame from each beat as input for the next.
-
Reject and regenerate locally
When a beat fails, redo that beat rather than the whole sequence.
Prompt approach
- With reference supplied, describe motion and camera, not the whole subject.
- Keep one action and one camera move per shot.
- State duration explicitly.
- Keep lighting consistent across shots in the same scene.
Practical limits
- Consistency improves but does not become perfect — expect some drift.
- Complex scenes with multiple subjects remain difficult.
- Fast motion and extreme angles still degrade the subject.
- Text and fine detail remain unreliable, as with all current models.