Model releases & benchmarks
VGI-Bench: Probing Visual Intelligence in Video Generation Models
VGI-bench tests whether video-generation models can reason through changing visual scenes—not merely produce a plausible final frame.
Why it matters: A plausible-looking clip is not enough to establish reliable visual reasoning. The benchmark’s emphasis on valid evolving processes and calibrated difficulty makes those evaluation choices explicit, while its reported results show substantial room for improvement.