NeurIPS paper: video diffusion models encode physics internally even when outputs are wrong
katjahofmann · x · 2026-10-13
The NeurIPS Spotlight paper "The Invisible Hand of Physics" asks whether video diffusion models (VDMs) encode physics internally even when their generated samples get it wrong. Answer: they do—and on average their internal physics understanding beats V-JEPA 2. The result suggests generation errors don't mean the model lacks physical intuition; advised by Meta's Katja Hofmann and others.
More from Multimodal
- Grok Bot turns any topic into cinematic explainer videos with ElevenLabs narration — XFreeze · 2026-10-13
- AI-animated kids' pilot The Brass Acorn pairs Pixar-style visuals with an original song — Warblewick · 2026-10-13
- AI-generated image turns Wikipedia's puzzle globe into a real soccer ball — Delahuntagram · 2026-10-13
- AI-generated 'knee pad ring' shrinks sports gear onto your finger — Delahuntagram · 2026-10-13
- AI-generated 'crypto heels' turn a bull-market candlestick chart into a shoe heel — Delahuntagram · 2026-10-13
- Shared: a 'cursed cam' prompt trick for Seedance 2.5 video generation — techhalla · 2026-10-13