Datasets:
The dataset viewer is not available for this split.
Error code: JobManagerCrashedError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
UnrealBokeh
UnrealBokeh is the synthetic dataset used by AnyBokeh (Project Page | arXiv) (NeurIPS 2026). It contains 11,477 images from 78 camera views in two Unreal Engine environments. The images, camera parameters, and ground-truth metric depth maps come directly from Unreal Engine. The disparity and signed circle-of-confusion (CoC) maps are produced by preprocessing and are included for convenience.
Files
Two environments, scene_00 and scene_01, contain 78 fixed camera views. A view keeps one position and several focal lengths. Each focal length has one metric depth map and one disparity map. Changing the aperture or the focus distance produces a new RGB image and a new signed CoC map, so the 11,477 images share 221 depth maps and 221 disparity maps. Every image is 2016 × 1120, and the camera sensor is 36mm × 20mm.
The paths follow this pattern:
data/<scene>/<view>/rgb/<view>_f<focal>_ap<aperture>_fd<focus>.jpeg
data/<scene>/<view>/coc/<view>_f<focal>_ap<aperture>_fd<focus>.npy
data/<scene>/<view>/depth/<view>_f<focal>.exr
data/<scene>/<view>/disparity/<view>_f<focal>.npy
The train split has 10,212 images from 71 views, and the validation split has 1,265 images from the other 7 views. No view appears in both splits. The files are packed into data/train_part01.zip, data/train_part02.zip, and data/validation.zip. Unzipping them restores the paths above, which are the paths used in data_lists/.
data_lists/ is the sampling used by AnyBokeh released checkpoints. A stage-1 entry is one image. A stage-2 entry pairs a source image with a target image from the same view.
Download
huggingface-cli download itsmag11/UnrealBokeh --repo-type dataset --local-dir UnrealBokeh
cd UnrealBokeh
for f in data/*.zip; do unzip "$f"; done
Preprocessing
RGB images, depth maps, and camera metadata come directly from Unreal Engine. preprocessing/generate_maps.py derives the disparity and signed CoC maps from those depth maps. It reads the focal length, aperture, and focus distance from the image filename. The sensor width is 36 mm.
pip install -r preprocessing/requirements.txt
python preprocessing/generate_maps.py all
Acknowledgement
The two environments are Military Airport and Research Tent Environment, created by FreshCan. FreshCan has permitted this release for academic research.
License
This dataset is released for academic research. Please cite the paper below when using the dataset. The environment assets remain the property of FreshCan.
Citation
@inproceedings{hou2026anybokeh,
title = {AnyBokeh: Physics-Guided Any-to-Any Bokeh Editing with Optical Fingerprint Transfer},
author = {Hou, Xinyu and Li, Xiaoming and Yue, Zongsheng and Loy, Chen Change},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2026}
}
- Downloads last month
- 224
