Interactive Generative Motion Editing via Scheduled Inpainting

Interactive motion editing has long sat in an awkward middle ground: traditional tools can nudge timing or pose, but larger structural changes often break the motion, while generative systems can create convincing new movement but struggle to preserve what was already there. Scheduled inpainting is designed to bridge that gap by balancing preservation and generation during inference, giving artists finer control over which parts of a motion stay fixed and which parts are regenerated.

The method works with generative models that already support direct manipulation, then layers on a scheduling approach that controls how aggressively the system inpaints over time and space. In practical terms, that means an animator can refine an existing clip instead of replacing it wholesale, while still keeping the result aligned with the learned motion distribution so it doesn’t drift into uncanny or physically inconsistent poses.

The most obvious game-dev use cases are motion extension, clip stitching, and compositing, which are exactly the kinds of tasks that eat time in cinematic pipelines and iterative animation passes. If it holds up outside demos, this could reduce the amount of hand-keyed cleanup needed when a shot changes late, a traversal needs to be lengthened, or two motion sources need to be blended into one seamless sequence.

The work was submitted to arXiv on July 31, 2026 by Dhruv Agrawal, Dominik Borer, Luca Vögeli, Robert Sumner, Martin Guay, and Jakob Buhmann. The team says it was validated against four baselines, with ablations and user feedback included, which...