Video models' physics failures traced to how positions are encoded

single source· 1 articles · confidence: medium · first seen 2026-09-19 20:00 UTC

What this means for you

Nothing to act on yet. This is a preprint with no code or weights offered, and the authors report improvements without numbers. If you serve video generation, the mechanism is the part worth knowing: a per-step frequency schedule for positional encoding, which the authors say can be tried without retraining.

Video generators often produce clips that break physical law. A preprint posted to arXiv on 19 September (2609.23658) looks inside text-to-video diffusion models and finds the motion trajectory is set in the earliest denoising steps, while the model is still turning noise into a rough shape. There, rotary position embeddings (which record where each item sits by rotating its vector) decay spatial attention too fast, so the model commits early to an implausible trajectory and suppresses plausible ones nearby. Scaling the embedding's frequency across denoising steps, the authors report, improves physical plausibility in experiments run both with and without retraining. No scores are given.

Key facts

  • ·Preprint arXiv 2609.23658 on text-to-video diffusion models, posted 19 September 2026. source
  • ·The paper identifies a specific subset of attention heads involved in the models' motion planning. source
  • ·Motion trajectories are reported to form during the early denoising stages, before detail is added. source
  • ·The authors attribute physically implausible output to excessive spatial attention decay caused by rotary position embeddings (RoPE). source
  • ·Proposed fix: scale RoPE frequency across denoising steps, described as a lightweight architectural modification. source
  • ·The paper reports training-free and training-based experiments improving physical plausibility, with no numerical results given. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire