LavalObjaverseDataset

vLAR Group | SIGGRAPH Asia 2026

A large-scale, high-quality dataset for multi-view relighting.

Dataset Summary

The Laval Objaverse Dataset is a comprehensive dataset designed for multi-view relighting and novel view synthesis tasks. It combines high-quality 3D assets from Objaverse with realistic, diverse illumination conditions from the Laval Indoor and Outdoor HDR datasets. Each render includes synchronized multi-view images, depth maps, and complete lighting metadata.

Dataset Construction

We structured our dataset through a rigorous four-step pipeline:

1. Object Filtering:
   We source base meshes from the Objaverse dataset. To ensure high visual fidelity, we exclude meshes with poor geometry or materials by adopting the strict object selection criteria from the relitObjaverse dataset [1].

2. Lighting Selection:
   We employ open-source illumination datasets, specifically the Laval Indoor HDR Dataset [2] and the Laval Outdoor HDR Dataset [3]. Each illumination map is uniformly rotated 16 times to maximize lighting diversity. For every object, we sample 16 environment maps with a balanced 1:1 ratio of indoor and outdoor lighting conditions.

3. Rendering Protocol:
   All objects are normalized to unit size (largest bounding box dimension = 1) and placed at the origin. Cameras are uniformly distributed on a sphere at distances of 1.8 or 2.2 units from the center, oriented toward the scene center.

4. Dataset Partitioning:
   The dataset is split into training, validation, and test sets with strict disjointness: no camera poses, illumination conditions, or 3D objects overlap between splits. The division information are under info/laval, info/objaverse and info/view for illuminations, objects and viewpoints respecitively.

Dataset Structure

LavalObjaverseDataset/
└── info/
│   ├── laval/
│   │   └── full_testing_lighting.json
│   │   └── full_training_lighting.json
│   │   └── full_validation_lighting.json
│   ├── objaverse/
│   │   └── all_objects.json
│   │   └── full_{split}_objects.json # split = testing, training, validation
│   ├── view/
│   │   └── all_views.json
│   │   └── full_{split}_views.json # split = testing, training, validation
└── pairs/
│   ├── 1_to_1_mapping_pairs.json
│   ├── 16_to_16_mapping_pairs.json
│   ├── 32_to_32_mapping_pairs.json
│   ├── olatverse.json
│   └── ...
│
└── rendered/
    ├── training/
    │   └── subset_{subset_uid}/
    │       └── {object_uid_0}.tar.gz
    │       │    ├── info.json
    │       │    ├── V18-Indoor_9C4A5417-f5d395fb63_6_image.png
    │       │    ├── V18-Outdoor_9C4A0573-f5d395fb63_6_image.png
    │       │    ├── V18_depth_0001.exr
    │       │    ├── V23-Indoor_9C4A5417-f5d395fb63_6_image.png
    │       │    └── ...
    │       └── ...
    │
    ├── testing/
    │   ├── {object_uid_1}.tar.gz
    │   ├── {object_uid_2}.tar.gz
    │   └── ...
    │   
    └── validation/
        ├── {object_uid_3}.tar.gz
        ├── {object_uid_4}.tar.gz
        └── ...

Inside Each Object Folder / Unzipped Archive:

• info.json: Metadata containing object information, rendering parameters, camera poses, and lighting conditions.
• Vxxx-yyy_image.png: Rendered RGB images where:
• Vxxx indicates the camera pose/viewpoint (e.g., V18, V23).
• yyy is the lighting map name (e.g., Indoor_9C4A5417-f5d395fb63_6 or Outdoor_9C4A0573-f5d395fb63_6).
• Vxxx_depth_*.exr: Corresponding ground-truth depth maps in OpenEXR format for each viewpoint.

File Naming Convention Example:

V18-Indoor_9C4A5417-f5d395fb63_6_image.png
│  │       │                   │
│  │       │                   └─ Rotating number (0,1,...,15)
│  │       │
│  │       └─ Lighting map identifier
│  └─ Lighting type (Indoor/Outdoor)
└─ Viewpoint ID (V18, V23, etc.)

References

If you use this dataset in your research, please cite our work along with the prior work we built upon:

@misc{wang2026relightformerfeedforwardgenerativetransformer,
      title={RelightFormer: Feed-forward Generative Transformer for Multiview Object Relighting}, 
      author={Hejun Wang and Jinxi Li and Junwei Jiang and Shiwei Mao and Hu Cheng and Shouwang Huang and Bo Yang},
      year={2026},
      eprint={2609.07414},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.07414}, 
}

@article{Jin2024,
    author = {Jin, Haian and Li, Yuan and Luan, Fujun and Xiangli, Yuanbo and Bi, Sai and Zhang, Kai and Xu, Zexiang and Sun, Jin and Snavely, Noah},
    journal = {NeurIPS},
    title = {{Neural Gaffer: Relighting Any Object via Diffusion}},
    year = {2024}
}

@article{laval_indoor,
    author = {Marc-Andr\'{e} Gardner and Kalyan Sunkavalli and Ersin Yumer and Xiaohui Shen and Emiliano Gambaretto and Christian Gagn\'{e} and Jean-Fran\c{c}ois Lalonde},
    title = {Learning to Predict Indoor Illumination from a Single Image},
    journal = {SIGGRAPH Asia},
    year = {2017}
}

@INPROCEEDINGS{laval_outdoor, 
    author={Hold-Geoffroy, Yannick and Athawale, Akshaya and Lalonde, Jean-François}, 
    booktitle={CVPR},  
    title={Deep Sky Modeling for Single Image Outdoor Lighting Estimation},  
    year={2019}
} 

Contact

For questions about the dataset, please contact:

• {hejun.wang, jinxi.li}@connect.polyu.hk
• [email protected]

For rendering code, models, and further details, please visit our GitHub Repository.