PRISM-Bench Evaluates Audio-Visual Coherence in T2AV Generation
September 7, 2026
PRISM-Bench is a diagnostic benchmark for text-to-audio-video generation, featuring 900 human-verified samples. It factorizes evaluation across audio types and sound-source visibility to assess audio-visual coherence and prompt following.
HOW THIS AFFECTS YOU
●
builderYou can use this to evaluate how well your T2AV models ground audio to on-screen visual events.
●
researcherThis provides a more rigorous way to diagnose failures in the audio modality of video generation models.