Why descriptions alone drift
Image and video models sample a new result every time. A line such as "a woman in her thirties with short dark hair" fits thousands of people, and the model fills the gaps differently on each run: eye spacing, jawline, hairline, height against the product. One image on its own looks fine. Put six side by side in a carousel, or cut four shots together in a film, and viewers notice that the person keeps changing. For a mascot or a recurring brand character, that drift quietly erodes the recognition the brand has paid to build.
Building a reference set
Start with a neutral front view, a three-quarter view and a profile, all in even light. Add a full-length image that shows proportions against something of known size, a small range of expressions, and each wardrobe item as its own reference. Write down the small identifying details, such as a scar, freckles, a ring or a particular parting, because models drop small details first. For illustrated characters, add a turnaround sheet, line weight, the palette as colour values, and a note of what may change by season and what never changes.
Keeping identity across video
Motion adds new ways to lose a face. Identity tends to slip in profile turns, during fast movement and whenever the character is small in frame. Establish the face clearly in the opening frames, prefer starting each clip from an approved still, and keep the character large enough in frame for the model to hold the features. Between shots, check wardrobe, hair, props and the direction of the light. In Synthetic White, characters are saved as Elements, so every image and clip is generated from the same approved references rather than from memory of a description.
Real people, likeness and consent
When a character is based on a real person, such as a model, an ambassador or an employee, consent comes first. It should name AI generation specifically and set out the channels, markets and duration, who approves outputs and what happens when the contract ends. Voice is a separate consent from face. Some places now regulate digital replicas directly: Tennessee's ELVIS Act of 2024 protects voice as well as likeness, and California's AB 2602 makes contract clauses unenforceable when they let a digital replica stand in for a person's own work without a reasonably specific description of the intended uses and the person had no lawyer or union representing them.
Updated 26 October 2026