NVIDIA's PhyProbe scores physical consistency of AI-generated videos, beating prior evaluators
nvidia · hf · 2026-10-14
NVIDIA, University of Waterloo, and Vector Institute introduce PhyProbe, an evaluator that maps features from a frozen spatio-temporal encoder to a scalar physical-consistency violation score via a lightweight head, trained with a unified objective combining pairwise ranking, regression on noisy annotations, and anchor-based calibration. It outperforms prior evaluators on most pairwise benchmarks, with the largest gains in no-correspondence and generated-generated settings where fine-tuned evaluators degrade sharply, correlates strongly with human judgments on a stable [0,1] scale, and remains competitive on human preference benchmarks despite never training on general preference labels.
More from Multimodal
- AI-generated NERDVANA tee mashes Nirvana's smiley with a dead Finder icon — Delahuntagram · 2026-10-14
- Alibaba's Representation Guidance: plain pixel-diffusion Transformers beat bigger models — bdsqlsz · 2026-10-14
- The Prompting Zone Debuts Episode 01: The Last Human — Inevitable_Emu2722 · 2026-10-14
- Narrative AI video on a single RTX 3090: 720p Wan2GP workflow with 2x2 storyboard references — Inevitable_Emu2722 · 2026-10-14
- 95% of new Chinese short dramas are AI-made; the format behind +780% ROAS Meta ads — CurieuxExplorer · 2026-10-14
- Singapore startup uses AI to turn novels into graphic novels, reaching 60-70% before humans finish — lmoroney · 2026-10-14