Why video models break physics: trajectory locked in first 10% of denoising
linoy_tsaban · x · 2026-09-25
Arguing that video models are now convincing enough that we should scrutinize their physical faithfulness, the author highlights a mechanistic interpretability study: an object's trajectory gets locked in during the first 10% of denoising, and RoPE's spatial decay locks it onto the wrong position. A one-line change to the rotary embedding fixes it. Since the paper only tested Wan2.1-1.3B, she is now applying the fix to H3 to see if it generalizes.
More from Multimodal
- Turn any song into a music video with Seedance 2.0: a two-step workflow — techhalla · 2026-09-25
- Surreal art showcase made with Recraft V4.1 Flash — aziz4ai · 2026-09-25
- Eerie AI-generated short film 'Morgue Cat' by Absurd makes the rounds — Able-Big9639 · 2026-09-25
- Segmented long video with H3 keeps degrading quality by the third segment — Delphoi_Studio · 2026-09-25
- Creative short film 'Cocinando con Energia' made with MiniMax H3 — Ok-Nerve941 · 2026-09-25
- Building a conceptual X-16 engine in 3D with GPT-6 Astra and Three.js — techartist_ · 2026-09-25