Video generator takes mid-generation velocity nudges instead of fixed positions
single source· 1 articles · confidence: medium · first seen 2026-09-14 20:00 UTC
What this means for you
Nothing to act on. There is no code, no weights and no evaluation date, so the 33% and 12% figures cannot be checked, and the preference figure comes from the paper's own human comparisons. Scope is tabletop rigid-body scenes only, well short of general video control.
A paper posted to arXiv on 14 September describes PhysStream, a video model that can be redirected mid-generation, not only before it starts. Existing controllable systems fix the motion schedule up front or steer by pixel position. PhysStream takes sparse velocity increments instead — how sharply an object's speed changes — and remembers position and tracking maps built from frames it has already produced, for tabletop scenes of rigid objects. The authors report a 33% cut in motion distribution distance and 12% lower trajectory error than their strongest baselines on synthetic benchmarks, and human preference in over 85% of comparisons. No evaluation date, code or weights are given.
Key facts
- ·The paper is arXiv 2609.17521, posted 14 September 2026. source
- ·Control is supplied as sparse velocity-increment signals rather than pixel-space object positions, so the full control schedule is not required before generation starts. source
- ·Structured scene memory is derived online from previously generated frames, using positional maps and object tracking maps. source
- ·Reported against the strongest baselines on synthetic benchmarks: a 33% reduction in motion distribution distance (FVMD) and 12% lower trajectory error. source
- ·Human evaluators preferred the model in over 85% of in-the-wild comparisons, according to the paper. source
- ·Evaluation is limited to multi-object tabletop rigid-body scenes; no code, weights or evaluation date is stated in the abstract. source
What the sources say
- Hugging Face Daily Papers (research) — Paper introducing the model, its two-stage training and its benchmark numbers
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersPhysStream: Streaming Physics-Grounded Video Generation with Structured Scene Memory and Fine-Grained Motion Control2026-09-14