Towards Unified Evaluation of Prompt Enhancers for Video Generation
PEBench is a new unified benchmark for directly evaluating video prompt enhancers across text-to-video, image-to-video, and reference-to-video tasks. It uses 1,100 expert cases and 35 fine-grained tasks to assess enhanced prompts based on semantic fidelity, internal consistency, and directorial creation. The evaluation shows that caption reconstruction and forward refinement methods have complementary strengths.