Physis-AI/wm-eval-gt-ewmbench

WM-Eval GT: EWMBench Ground Truth

Ground truth data for EWMBench evaluation, extracted and restructured for direct use by the wm-evaluation-harness framework.

Source

Extracted from agibot-world/EWMBench (gt_dataset.tar). Original dataset by AgibotTech.

Dataset Structure

{task_id}/
├── {episode_id}/
│   ├── prompt/
│   │   ├── init_frame.png      # Initial frame for generation
│   │   └── prompt.txt          # Task description
│   └── video/
│       ├── frame_00000.jpg     # Ground truth frame sequence
│       ├── frame_00001.jpg
│       └── ...

Statistics

Item	Count
Tasks	7 (367, 392, 497, 511, 543, 558, 574)
Episodes per task	3
Total episodes	21
Total files	3,959
Size	~129 MB

Usage with wm-evaluation-harness

This dataset is auto-downloaded by the framework when referenced as a HuggingFace dataset repo ID:

# configs/ewmbench_hf.yaml
evaluation:
  gt_dir: Physis-AI/wm-eval-gt-ewmbench    # auto-download
  benchmarks:
    ewmbench:
      metrics: [psnr, ssim]

wm-eval evaluate --config configs/ewmbench_hf.yaml --videos <your_pred_dir>

Or used programmatically:

from wm_eval.evaluation.models.registry import ModelRegistry

registry = ModelRegistry()
gt_path = registry.resolve("Physis-AI/wm-eval-gt-ewmbench", repo_type="dataset")

License

This dataset follows the original CC BY-NC-SA 4.0 license from AgibotTech/EWMBench.

Citation

@article{ewmbench2025,
  title={EWMBench: Evaluating Scene Understanding and Generation Quality for Embodied World Models},
  author={AgibotTech},
  year={2025},
  eprint={2505.09694},
  archivePrefix={arXiv}
}