WRBench

WRBench is an open benchmark for camera-controlled generation and diagnostic evaluation of video world models, introduced in Current World Models Lack a Persistent State Core.

Links

• Paper: https://huggingface.co/papers/2606.20545
• Project page: https://jinplu.github.io/WRBench/
• GitHub: https://github.com/JinPLu/WRBench
• Artifact collection: https://huggingface.co/collections/WRBench/wrbench-current-world-models-lack-a-persistent-state-core-6a365c717251293c9fc2cc26
• Leaderboard: https://huggingface.co/spaces/WRBench/wrbench-leaderboard

Datasets

• Natural-25 prompts, variants, and first frames: https://huggingface.co/datasets/WRBench/wrbench-natural25
• 23-model evaluation results: https://huggingface.co/datasets/WRBench/wrbench-results
• Human annotation verdicts: https://huggingface.co/datasets/WRBench/wrbench-human-annotations
• Benchmark videos and per-video scores: https://huggingface.co/datasets/WRBench/wrbench-videos