What are the challenges in evaluating generative media models?

Existing benchmarks often focus on isolated modalities or coarse similarity, failing to capture the fine-grained joint correctness required for realistic text-to-audio-video generation.