Character Persistence and Motion Transfer

Multi-shot AI generation faces a fundamental consistency problem: characters change appearance between shots, and transferring specific motion from one sequence to another remains imprecise. We present two complementary approaches — Cast for reference-based identity persistence and Motion for choreography transfer — that address the inter-shot coherence gap in professional generative workflows. Together, these systems enable directors to maintain character identity while transferring performance across shots.

When generating multiple shots of the same character, current AI systems treat each generation as independent. The result: facial features drift, clothing changes, hair color shifts, and body proportions fluctuate. For single-image use cases, this is tolerable. For multi-shot pre-visualization — where a character must be recognizable across an entire sequence — it is a production blocker.

This problem is distinct from the broader temporal coherence challenge. While temporal coherence addresses environmental and stylistic consistency, character persistence specifically targets the preservation of individual identity — the features that make a character recognizable as the same person across shots.

The Cast approach uses reference images to anchor character identity across generations. A creator provides one or more reference images of a character, and the system extracts identity parameters — facial structure, skin tone, hair characteristics, body proportions — that persist across all subsequent generations featuring that character.

• Facial structure anchoring: Bone structure, facial proportions, and distinguishing features are parameterized and locked.
• Costume persistence: Clothing, accessories, and wardrobe details are maintained across shots within a scene.
• Multi-angle consistency: The character remains recognizable whether shot from the front, profile, or three-quarter angle.

Motion transfer addresses a different challenge: applying specific movement, gesture, or performance from one video to a generated sequence. A director might want a character to perform a specific walk, gesture, or action that they've captured in reference footage. The Motion approach extracts the temporal movement data from the reference and applies it to the generated character — preserving the choreography while allowing the visual treatment to match the project's art direction.

Character persistence extends our Continuity Lock research (documented separately) into the video generation domain. While Continuity Lock addresses identity preservation in still-frame sequences within the Keyframe workspace, Cast and Motion address the additional challenges introduced by temporal generation — movement, expression changes, and dynamic lighting interactions that occur across video frames. Together, these systems provide comprehensive identity management across the entire AIM Previz pipeline.

Character persistence and choreography transfer are complementary capabilities that address the two fundamental questions in multi-shot AI generation: "Does this look like the same person?" and "Does this capture the performance I want?" By solving both, we enable AI-generated pre-visualization sequences that maintain the narrative coherence professional production demands.