Interactive Generative Motion Editing via Scheduled Inpainting

How to read this

Category

research paper, inference time technique for interactively editing existing motion clips with pretrained generative models

Contributions

• Introduces scheduled inpainting, an inference only technique requiring no additional training, that extends pretrained spatially conditioned generative motion models with the ability to edit existing motion clips rather than only generate new ones
• Combines a user controlled temporal schedule with a spatiotemporal strength mask, giving artists fine grained control over where and how much original motion is preserved versus regenerated
• Unifies motion synthesis and editing in a single framework, supporting extending, stitching, and compositing of different motion clips along with direct point dragging edits, all in the same real time system
• Validates the approach against four baselines including DDIM noise inversion and noise optimization, with ablations of the design and reported feedback from professional artists

Context

This paper builds directly on the cited Generative Motion Rig for Artist Driven Motion Authoring (Buhmann et al, 2026), applying inference time control on top of that kind of pretrained spatially conditioned diffusion model, and specifically on the direct manipulation motion models IBMM and SF control that it operates with. It positions itself against text conditioned motion editing systems like MotionFix and MotionLab, arguing text is too coarse a control modality for precise artist edits, and against naive inpainting and blending techniques, which the authors show leave inpainted regions completely uneditable.Builds on: A Generative Motion Rig for Artist-Driven Motion Authoring

Correctness

The paper reports comparison against four adapted baselines and ablations of its own design choices, the schedule, the spatiotemporal mask, and the normalization state needed to composite clips from different global orientations and scales, plus qualitative feedback from professional artists. It is upfront that naive inpainting alone is not sufficient, which is the actual technical justification for the more elaborate scheduled approach, though this guide reviewed the introduction and method rather than the full quantitative results tables.

Clarity

Readable for someone with prior exposure to diffusion based motion generation and inpainting terminology from the image domain, DDIM, latent noise inversion, and the paper explains its motivating problem, small edits fine, big structural edits break, in plain, production relevant language before the technical formulation arrives.

How to read it

Five minute pass: read the abstract and Figure 1's stitched, extended, and dragged clip examples to see the range of edits scheduled inpainting supports. Second pass: read the introduction's three contributions and Section 3's inpainting schedule formulation to see how the timestep regulated scalar and spatiotemporal mask combine to control preservation strength. Third pass is worth it for anyone building artist facing motion editing tools on top of a pretrained diffusion motion model, since the normalization state detail for compositing clips from different global orientations is a small but easy to miss implementation point.