This dataset is derived from DL3DV-10K and is shared for non-commercial research only. By requesting access you confirm that you have accepted the DL3DV-10K terms of use (https://github.com/DL3DV-10K/Dataset), that you will use the data under the same terms, and that you will cite GenWildSplat (https://genwildsplat.github.io/) and DL3DV-10K in any work that uses it.

GenWildSplat Data: relit outdoor DL3DV-10K scenes

Training data of GenWildSplat — Generalizable Sparse-View 3D Reconstruction from Unconstrained Images (CVPR 2026, project page, arXiv). It contains 714 outdoor scenes from DL3DV-10K whose frames were relit under random lighting with DiffusionRenderer, together with the original DL3DV camera poses.

Data processing

Scene selection. We selected outdoor scenes from the 1K and 2K batches of DL3DV-10K (359 + 355 = 714 scenes).
Relighting. We used DiffusionRenderer (NVIDIA; Liang et al., CVPR 2025) to relight each scene's frames at 1/8 resolution under randomly sampled lighting, so the images show the same scenes under different illumination than the original DL3DV captures.
Cameras. Camera intrinsics and extrinsics are DL3DV's original transforms.json; they are unchanged by relighting.

Only the relit frames and the cameras are released; DiffusionRenderer's intermediate maps are not included.

item	description
`1K/<scene>.tar`, `2K/<scene>.tar`	one archive per scene (359 from DL3DV's `1K` batch, 355 from `2K`); the scene name is DL3DV's 64-character hash
`<scene>/images_8/frame_XXXXX.png`	the scene's relit frames at 1/8 resolution (480 × 270), typically 330 per scene
`<scene>/transforms.json`	camera intrinsics and per-frame extrinsics in nerfstudio format, from DL3DV. The intrinsics (`fl_x`, `fl_y`, `cx`, `cy`, `w`, `h`) refer to the full 3840 × 2160 resolution; divide them by 8 for `images_8`.
`train/index.json`	the 714 scenes used for training (`"1K/<hash>"`, ...)
`test/index.json`	the 10 scenes used for testing

Loading

from huggingface_hub import snapshot_download
import tarfile, pathlib

root = pathlib.Path(snapshot_download("vinayakgupta/GenWildSplat-Data", repo_type="dataset"))
for tar in root.glob("*K/*.tar"):
    with tarfile.open(tar) as t:
        t.extractall(tar.parent)          # -> 1K/<hash>/images_8, 1K/<hash>/transforms.json

License

The data is derived from DL3DV-10K and remains under its license: CC BY-NC 4.0 together with the DL3DV-10K Terms of Use (non-commercial research and education only; the data may only be shared with researchers who have themselves agreed to those terms). Access is therefore gated and approved by the owner. The relit images were generated with DiffusionRenderer; its own license terms also apply.

Citation

If you use this data, please cite both GenWildSplat and DL3DV-10K:

@article{gupta2026genwildsplat,
  title   = {Generalizable Sparse-View 3D Reconstruction from Unconstrained Images},
  author  = {Gupta, Vinayak and Lin, Chih-Hao and Wang, Shenlong and Bhattad, Anand and Huang, Jia-Bin},
  journal = {CVPR},
  year    = {2026}
}

@inproceedings{ling2024dl3dv,
  title     = {Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision},
  author    = {Ling, Lu and Sheng, Yichen and Tu, Zhi and Zhao, Wentian and Xin, Cheng and Wan, Kun and Yu, Lantao and Guo, Qianyu and Yu, Zixun and Lu, Yawen and others},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages     = {22160--22169},
  year      = {2024}
}

Downloads last month: 21

Total file size:

48.8 GB

Paper for vinayakgupta/GenWildSplat-Data

Generalizable Sparse-View 3D Reconstruction from Unconstrained Images

Paper • 2604.28193 • Published Apr 30

You need to agree to share your contact information to access this dataset