FilmBench (Reference-to-Video): leaderboard
FilmBench (2026) scores video-generation models on prompts reverse-engineered from clips of award-winning films in 20 genres, most scripting several shots, using an automatic evaluator built with film-academy directors on a three-level cinematic-language taxonomy. This board covers the 654 reference-to-video prompts (text prompts with a reference scene, character or prop image); score is the equal-weighted mean of the instruction-following, temporal-continuity and aesthetic-quality axes on a 0-100 scale, higher is better.
Metric: Overall score (0-100), the equal-weighted mean of the three axes, on 654 reference-to-video prompts; prompts reverse-engineered from clips of award-winning films in 20 genres, scored by the authors' expert-grade evaluation agent (Cinematic Language operators plus a judge model) on a three-level cinematic taxonomy, 1-5 per sub-metric mapped to 0-100. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Seedance 2.0 | 86.7 | |
| 2 | HappyHorse 1.1 | 85.5 | |
| 3 | HappyHorse 1.0 | 84.9 | |
| 4 | Kling 3.0 Omni | 80.4 | |
| 5 | Veo 3.1 | 78.2 | |
| 6 | Grok Imagine Video | 76.9 | |
| 7 | Vidu Q2 Pro | 71.4 |
Interactive version: theaggregate.ai/benchmark?slug=filmbench-reference-to-video · How It Works · Data refreshed daily, snapshot 2026-11-13.