The hardest problem in AI video is keeping the same person or product looking the same in every shot. Wavemaker solves it with a subject reference system that runs automatically on every generation.
How it works
- Detection. Anything that recurs — a character, mascot, product, logo, vehicle — is detected at planning time. Anything appearing in two or more scenes gets a reference.
- A canonical reference is built. From your uploaded photo, a scraped product image, a found photo of a named public figure, or a clean synthesized design. References pass their own quality review (a deformed or off-model reference is rejected and rebuilt).
- Multi-angle coverage. Characters and products get additional synthesized viewing angles, so a three-quarter shot doesn’t drift from the front view.
- Injection everywhere. The reference (plus angles) rides into every scene that features the subject — both the still frame and the motion generation — locking identity end-to-end.
- Verification. A consistency review compares scenes and flags identity, wardrobe, or style drift for automatic retakes.
What’s kept consistent
- Faces & bodies — the same person, every scene. For real people (your upload or a public figure), the face is locked while wardrobe follows the script — so the character can change outfits when the story calls for it, but never accidentally.
- Wardrobe — a recurring character keeps one canonical outfit across scenes unless the script deliberately changes it.
- Products — shape, materials, colors, and label text. Brand text on a product is transcribed from your real photo and spelled letter-perfectly in generated scenes.
- Logos — never re-rendered; your acquired logo file is composited with exact pixels.
Getting the best identity lock
- Upload a real photo. Your uploads are trusted as ground truth — a real photo beats any synthesized reference.
- One clean, front-facing, well-lit photo of a person; a straight-on, uncluttered photo of a product.
- Name your subjects. “Maya, the barista” is trackable across the script; “a woman” in five scenes may be five castings.
- Describe the look you want once (“tall, red hair, 30s”) — it binds the casting everywhere.
Brand kits
Your organization can have a saved brand kit — logo, colors, brand voice — that layers on top of anything scraped from a URL: every generation for your org starts from your locked identity, site research fills the gaps, and the kit wins on conflict. Brand kits are currently set up per organization — contact us to enable one for your team.
Subjects persist across edits
Subject references are saved with the composition. When you come back tomorrow and say “add a scene of Maya waving goodbye”, it’s the same Maya — same face, same outfit, same voice.