Keep an AI video character consistent across shots

Decide what must stay the same

Before generating coverage, write a short character description and choose a small set of approved references. Separate fixed details from things that can change with the shot.

For example, Evelyn enters an archive wearing a dark coat. Her next shot is a close-up as she turns towards a light. The camera and lighting change; her face, hair, and costume should remain recognisable.

Keep references that clearly show those details. A beautiful wide shot may be a poor face reference if the character occupies only a few pixels. Conflicting costumes or hairstyles make the intended identity harder to communicate.

Work through a three-shot scene

1. Establish the character

Generate or choose a medium shot of Evelyn entering the archive. Check her appearance against your reference pack before approving it. Keep the prompt, supplied references, model, and settings with the output.

Write down why this take works: “Face matches the approved portrait; coat collar is correct; empty hands.” These notes become checks for the next shot.

2. Change the framing, then review

For the close-up, reuse the approved character references through the next model’s supported reference controls. Describe the new framing and action while keeping the identity and costume description consistent.

Compare the result beside the first shot. Check the face and hair first, then costume and scene state. If the coat changes, address that mismatch before adding camera movement or another action. Changing several things together makes it harder to tell what helped.

3. Check continuity in sequence

Place the approved shots together in the edit. A take that looks plausible alone can still produce a distracting change at the cut.

Review at normal playback speed, then pause around the transition. If the character no longer reads as the same person, return to the references and the generation record for that take. Keep the rejected attempt with a brief reason so you do not repeat it unknowingly.

Keep a record for each attempt

A simple naming and note convention helps even without a dedicated app:

• Shot: archive-020-close-up
• Attempt: take-03
• References: approved Evelyn portrait and dark-coat reference
• Generation: provider, model, prompt, and available settings
• Decision: selected or rejected, with the reason

A seed can be useful to record when the model exposes one, but it is not a portable character identity across models. A new provider may also accept different numbers or types of references. Check those controls instead of assuming a prompt alone carries everything forward.

Using Velvet Avocado

Velvet Avocado lets you organise character assets and inspect the recorded prompt, inputs, and settings beside generated media. Use that record as a starting point for the next attempt, and bring chosen media into the Editor to assess the sequence.

The benefit is being able to find and reuse your work across a project. It does not remove the need to judge likeness or overcome a model’s limitations.

When references are not enough

If repeated attempts still change the character, try a simpler shot or a model with reference controls better suited to the task. Some workflows use character-specific training, but that adds preparation and is not interchangeable across every generation service.

Keep the approved references and the review process either way. They give you a consistent basis for deciding whether a new take belongs in the scene.