AVBench: Human-Aligned and Automated Evaluation Benchmark for Audio-Video Generative Models
Researchers have introduced AVBench, a new automated benchmark designed to evaluate audio-video generative models, particularly those focused on human-centric scenarios. The benchmark incorporates fine-grained metrics across visual quality, audio quality, and cross-modal consistency, aiming to capture details often missed by existing evaluations. AVBench utilizes specialized evaluators trained through preference learning on a large dataset, deriving continuous scores from binary decisions to better align with human judgment and serve as a reward signal for RLHF. AI
IMPACT Provides a more accurate and automated method for assessing the capabilities of audio-video generative models.