NVIDIA's Physis-Lang Boosts Video World Models' Physics, Letting Cosmos3-Nano Surpass Veo 3.1
nvidia · hf · 2026-10-01
- Video world models often produce plausible-looking videos that violate basic physics. NVIDIA's Physis-Lang revisits the assumption that natural language can't represent physical knowledge, treating physics language as a shared, optimizable representation across data curation, training, and generation.
- Physical processes are described via entities, causes, interactions, governing principles, temporal evolution, and effects; PhysCapBench decomposes processes into atomic assertions scored by recall/precision; an agentic loop iteratively analyzes assertion-level errors and refines captioning instructions; model deficiencies are converted to text for language-guided retrieval of videos covering missing physics.
- Consistent physical-plausibility gains across four benchmarks on Wan and Cosmos backbones; starting from open-source Cosmos3-Nano, Physis-Lang-enhanced models surpass the leading proprietary Veo 3.1.
Related event: NVIDIA's Physis-Lang Tops Physics-IQ Video Benchmark(4 posts)→
More from Multimodal
- New ChatGPT image failure mode: faces warp on handstand poses — flowersslop · 2026-11-15
- Fable 5.2 composes a piano piece 'Elsewhere, Then Here' and animates its own video — Antique_Cat4404 · 2026-11-15
- Creator claims no AI video platform makes money: Sora shut down, Grok Imagine layoffs — AIandDesign · 2026-11-15
- Use handheld camera feel instead of steadicam to add realism to AI video scenes — moplee1990 · 2026-11-15
- Coding over image models: pure-code art generation in Gemini and the app's nano banana bias — andrew_n_carr · 2026-11-15
- Midjourney style code 2659353360 yields striking ink-splash Marvel art — azed_ai · 2026-11-15