As joint audio-video generation models improve, they can produce scenes with three forms of complexity: multi-event composition, multi-layer audio mixture, and asymmetric cross-modal relations. These go beyond simple one-to-one correspondences, making evaluation harder.

The paper introduces AVC-Reward, an event-centric approach to audio-video consistency evaluation. It is designed to handle these complex scenarios, offering a more granular assessment than existing metrics.

This matters because generation models are increasingly expected to create realistic, multi-layered scenes. A metric that explicitly accounts for event structure and asymmetric relations could better reflect perceptual quality and guide further model development.