PVBench: New Benchmark Scores Video Generation Across 46 Physical Plausibility Dimensions
Kaiwei Chen · hf · 2026-10-21
Researchers from Zhejiang University's APRIL Lab, HUST, and Tencent Youtu Lab released PVBench, a benchmark that quantitatively measures physical plausibility in AI-generated videos.
- Covers 46 dimensions across 7 domains: instruction/semantic fidelity, behavioral & causal consistency, perceptual & temporal quality, geometric/spatial consistency, material properties, kinematic persistence, and mechanical plausibility with conservation laws
- Built from 600 crafted prompts, 30% involving physically embodied agents (robot arms, humanoids, quadrupeds)
- Pairs hard-to-evaluate dimensions with specialized vision models extracting verifiable object-level measurements (6D pose, metric scale, surface normals, kinematics), moving beyond coarse VLM-based judgment
More from Multimodal
- Fibo2 old-photo restoration and colorization tested on a historic fire scene — Striking-Long-2960 · 2026-10-21
- Full Seedance 2.5 prompt shared for generating ultra-realistic handheld smartphone video — techhalla · 2026-10-21
- GPT Image 2.5 vs Nano Banana 2.1: 502 synthetic rust masks reviewed, 79.7% retained — Gloomy_Recognition_4 · 2026-10-21
- AI-generated 'Doner Kebab Taxi' video goes viral, poster holds the music rights — average1804 · 2026-10-21
- Fibo 2.0 ComfyUI workflows: Edit Turbo is fast and solid at image editing — nomadoor · 2026-10-21
- MOSS t2i anti-blur tip: write 'photography', not 'photograph' — Bulky_Chemistry_196 · 2026-10-21