FilmBench: A Film-Grade Benchmark for Cinematic Video Generation

Researchers introduced FilmBench, a benchmark built from award-winning film clips and professional cinematic language for video generation evaluation.

It tests both text-to-video and reference-to-video systems with a three-level metric taxonomy and an expert automatic evaluator.

The release also includes open-sourced FilmOps operators to support richer cinematic prompts and assessment.

Results showed strong agreement with human rankings, but leading models still scored low overall and struggled especially on multi-shot prompts.

FilmBench raises the bar for AI video evaluation by using film-industry criteria that reveal weaknesses missed by simpler web-style benchmarks.