Why your AI character looks different every shot.
The identity drift problem in generative video, and our N-1 reference approach for keeping a single character locked across an entire session.
If you've spent any time generating video with current foundation models, you've seen it: a character that's a different person in every shot. Same outfit, same brief, same prompt. Different face.
We call this the drift problem. The model isn't refusing to keep your character consistent. It just doesn't know what "your character" means past the prompt. It re-rolls the face every frame.
The N-1 reference
Our approach is to never let the model start from scratch. Every new scene is conditioned on the last known-good frame, plus a learned signature of the character. The model isn't generating a fresh person. It's continuing one.
That's what gets us 0.91 AuraFace similarity across 50+ scenes per session.
What this means for you
You lock a character to a reference image once. From there, every scene you render — close-up, side angle, outdoor, talking to camera — is the same character. Same face, same eyes, same presence.
More on the benchmark methodology in the next post.