The most useful motion papers this week are less about making a clip look spectacular and more about giving a generator a structure it can reason over. Three new preprints put physics, retrieval, and language at different points in the pipeline.

Physics before pixels

PhysFlow separates motion control from appearance. Its physics-aware optical-flow stage estimates foreground motion, while a second FlowRender stage turns that signal into video. The authors release PhysVideo 10K, with 10,000 foreground objects across 50,000 sequences. The design is a practical reminder that “motion control” is not the same as copying a reference frame: a useful control signal should say how an object moves, not merely where it was seen.

Retrieval for plausible motion

ReMoMask-2 adds a retrieval-augmented path to masked motion generation. A structure-aware database supplies both global and body-part cues; topology-aware attention keeps those cues connected during generation. On KIT-ML and SnapMoGen, the paper reports the lowest FID among its comparisons while retaining fast inference. Retrieval here is not a search result pasted onto a prompt. It is a way to constrain the latent space before the model fills in missing motion.

Motion as a language citizen

Open-UniMo treats motion understanding and generation as one open-world problem. Its million-scale motion-language corpus, 64K motion-token vocabulary, motion-consistent chain-of-thought training, and GRPO optimization are aimed at making motion a first-class sequence in a language model. The striking claim is bidirectional: generation improvements also help understanding on the authors’ Open-MoBench.

Together these papers suggest a stack: estimate a physically meaningful signal, retrieve structurally compatible examples, then let language coordinate the result. That stack is more legible than a single end-to-end score, and it gives practitioners clearer places to test failure.

For context, compare this week’s structured controls with our earlier coverage of motion-blind evaluation and sound-conditioned movement. The open question is whether those evaluation and multimodal layers can survive contact with real users, not just curated benchmarks.

References