Interactive Motion Editing That Regenerates While Preserving Existing Movement
Scheduled inpainting lets artists interactively balance preservation and regeneration in existing animation for natural motion extension, stitching, and compositing.
Paper Interactive Generative Motion Editing via Scheduled Inpainting
Research DisneyResearch|Studios · ETH Zurich
Release 2026-07-30
Combining motion synthesis and editing in a single workflow
Researchers at DisneyResearch|Studios have proposed “scheduled inpainting,” which preserves existing animation while regenerating only the necessary segments with a generative model. In VFX and game production, existing motion may need to be modified when the environment or direction changes, but traditional editing can cause distortions after major structural changes, requiring the motion to be created again from scratch.
Recent generative motion models can create entire movements from sparse spatial conditions, but they have limitations when it comes to interactively modifying existing motion while preserving it. Scheduled inpainting connects the two tasks by adjusting, during inference, the balance between areas where the original should be preserved and areas that should be newly generated according to time and space.
Artists can directly move parts of a character's trajectory or pose and repeatedly refine the results. The system produces natural results from the learned motion distribution while preserving unedited sections, and supports tasks such as extending and connecting clips and blending different movements.
The paper compares the approach with four baseline methods and validates it through ablation experiments and user feedback. The researchers are Dhruv Agrawal, Dominik Borer, Luca Vögeli, Robert W. Sumner, Martin Guay, and Jakob Buhmann. We correctly linked the post to the paper rather than treating it as a simple demonstration.