The authors constructed PVBench, a benchmark covering 46 evaluation dimensions organized into seven domains: instruction fidelity, behavioral consistency, perceptual quality, geometric consistency, material properties, kinematics, and mechanical conservation laws.
New benchmark quantifies physical plausibility in AI-generated videos
PVBench uses specialized vision models to measure 46 physical dimensions, revealing that current video generators consistently violate object-level physics.
Top University
Kaiwei Chen · Moran Li · Zhenye Gan · Yabiao Wang · Chengjie Wang · Kexin Zhang · +2 more
Zhejiang University
Research Digest··3 min read
Kaiwei Chen and colleagues introduce PVBench, a benchmark that systematically evaluates physical plausibility in text-to-video and image-to-video generation.
§