New Benchmark Tests Physical Commonsense in Audio-Video AI

A team of researchers has developed a new assessment framework called AV-Phys Bench to evaluate joint audio-video generation models and their grasp of real-world physics. This benchmark focuses on three primary components: Steady State, Event Transition, and Environment Transition, further subdivided into categories based on actual physical principles. It features challenging Anti-AV-Physics prompts that test the limits of models' capabilities. Evaluators consider five key criteria when analyzing outputs. Of the seven models tested, Seedance 2.0 emerged as the top performer, highlighting deficiencies in many models regarding true cross-modal physical coherence versus mere superficial believability.