What it is
A 57.7K-clip game-engine video dataset plus benchmarks for rendering structured controls like grey boxes, semantics, masks, and skeletons into finished video. ~57.7K training clips across 3 sources plus 2 benchmark splits. Each sample pairs a rendered video with structure video, first frame, and prompt; controls include grey boxes, semantic maps, masks, and skeletons.
Key results
- 57,738 training clips across Magpie, CARLA, and UnrealZoo
- 208 hard benchmark cases + 53 held-out clips
- Each clip has render.mp4, condition.mp4, first_frame.png, and json
Useful for training and benchmarking structure-conditioned video generation from engine or agent-produced controls.