Keep an AI avatar consistent across ad scenes
Build an identity reference pack, lock the details that matter and review continuity before assembling a multi-scene ad.
By Janice Phil · Growth · Published
In this article
Define what needs to stay consistent
Consistency means the viewer recognizes the same person and product across shots. It includes more than facial resemblance: hairstyle, wardrobe, voice, product scale, light direction and scene geography can all make a sequence feel disconnected.
Write down the details that must not change and the ones you intend to vary. A new camera angle is a deliberate change; a different neckline appearing halfway through the ad usually is not. Without that distinction, reviewers can disagree about whether a take is acceptable.
Build a reference pack for one identity
Choose one authorized avatar reference and keep it associated with the whole concept. Add useful alternate views only when the workflow supports them and their identity is consistent. Do not mix attractive portraits of different people and expect the generator to infer one stable character.
The three portraits used in this article’s cover are different demo avatars. They represent casting choices, not a claim of identity consistency across generated shots. The single reference below is the sort of asset to lock once you choose a performer.
Write a compact continuity sheet
Keep this alongside the script and output files. It is a review aid, not a guarantee that a model will preserve every property.
| Element | Lock | Allowed variation |
|---|---|---|
| Identity | Same approved face reference | Expression appropriate to the line |
| Wardrobe | Color, neckline and visible accessories | Natural folds |
| Product | SKU, packaging and dimensions | Camera angle |
| Setting | Background and light direction | Framing within the location |
| Voice | Approved speaker and pronunciation | Pace and emphasis |
| Shot | Screen direction and hand holding the item | Planned cut or camera move |
Approve an anchor scene first
Start with a simple shot where the face and product are easy to inspect. Once it meets the brief, use the same approved references for the next shot. Do not select an anchor only because its lighting is attractive if the product is already wrong.
In roasOS, Avatar mode exposes reference inputs and model-dependent controls. Use the selected model’s supported reference path. A saved avatar is helpful source material, but it does not remove the need to inspect each generation.
For speech continuity, keep the approved script and voice approach stable. Where a separate lip-sync workflow is used, check timing and pronunciation after each step; visual similarity alone will not make audio feel continuous.
Compare scenes before committing to the edit
Place the key frames next to one another and check face shape, hairline, clothing, hand position and product size. Then watch the cuts in sequence. A mismatch that seems minor in isolation can become obvious at the transition.
If a shot changes identity, regenerate or replace it rather than describing it as a stylistic variation. If only the angle is awkward, adjust the edit or use a cutaway. Keep rejection notes specific enough that the next attempt addresses the cause.
Store approved references and settings with the final sequence. This makes a future batch easier to reproduce and prevents a new teammate from unknowingly recasting the ad.