Gen3 scores highest on all four Video-Bench quality measures, and motion is every model's weakest

What the chart shows

Video-Bench uses GPT-4o as a judge, calibrated against human ratings, to score generated videos. Gen3 scored highest of the seven models on all four quality measures: imaging quality (4.66), aesthetic quality (4.44), temporal consistency (4.74) and motion effects (3.99). Kling had the second-highest average across the four (3.89), just ahead of Cogvideox (3.85).

Motion effects is the lowest of the four scores for every model in the data, from Gen3’s 3.99 down to LaVie’s 2.43.

Key numbers

How to read it

• Video-quality dimensions only; Video-Bench also scores how well videos match their prompts.
• The evaluator is a multimodal model (GPT-4o) with few-shot scoring; the report says its scores track human ratings more closely than earlier benchmarks.
• Averages across the four measures (Kling 3.89, Cogvideox 3.85) are our calculation; the report also ranks Kling second overall.