VisPhyWorld-Sub-All

# VisPhyWorld Sub All This dataset fuses the `sub` split ground-truth videos from **VisPhyBench / VisPhyWorld** with the corresponding generated videos from `TIGER-Lab/VisPhyWorld-Sub-Generated-Videos`. It is intended for sample-level inspection, model comparison, qualitative evaluation, and downstream tooling that benefits from having ground-truth and generated videos in one repository. The source benchmark collection is available at: - Dataset: `TIGER-Lab/VisPhyBench-Data` - Generated-video companion dataset: `TIGER-Lab/VisPhyWorld-Sub-Generated-Videos` - Collection: `TIGER-Lab/visphyworld` - Repository: `https://github.com/TIGER-AI-Lab/VisPhyWorld` ## Contents The repository is organized as: ```text GT/*.mp4 threejs/<model>/*.mp4 p5js/<model>/*.mp4 video/<model>/*.mp4 detection_json/*.json metadata.jsonl difficulty_table.json metadata.json ``` `GT/` contains a copy of the original `data/sub/videos` files with the same sample filenames. The generated-video folders preserve the engine/model hierarchy used by the companion generated-video dataset. ## Metadata - `metadata.jsonl` contains one row per benchmark sample in the `sub` split. - `video_path` points to the corresponding ground-truth file under `GT/`. - `detection_json_path` points to `detection_json/` when an annotation file is available. - `metadata.json` provides dataset-level counts, layout information, and source references. ## Notes - Filenames are aligned across ground-truth and generated videos wherever outputs are available. - 3D generated files use the remapped benchmark names, for example `task00001_3D_000.mp4`. - This repository is a fused convenience dataset for browsing and comparison; the original benchmark dataset remains hosted separately in `TIGER-Lab/VisPhyBench-Data`. ## Citation If you use this repository, please cite the VisPhyWorld / VisPhyBench project: ```bibtex @misc{visphyworld2026, title = {VisPhyWorld: Probing Physical Reasoning via Code-Driven Video Reconstruction}, author = {Liang, Jiarong and Ku, Max and Hui, Ka-Hei and Nie, Ping and Chen, Wenhu}, year = {2026}, eprint = {2602.13294}, archivePrefix = {arXiv}, primaryClass = {cs.CV} } ```