Identity-Preserving Video Generation

This field studies how to generate or edit video while keeping a referenced person or character recognizable across changes in pose, expression, viewpoint, clothing, and scene. Common approaches condition diffusion or video-transformer models on reference images or clips, using identity tokens, attention or lightweight adapters, multi-view cues, and separate controls for identity and motion. For longer videos and multiple subjects, systems also use keyframe anchors, entity memories, pose binding, or streaming caches to limit identity drift and confusion. A central challenge is balancing faithful identity with natural motion, prompt control, and visual variety, while evaluations still need to account for expression, occlusion, viewpoint, and consistency over time.