Omni-Primitive-Transforms

Omni-Primitive-Transforms is a 3D object primitive-transformation dataset. It can be used to generate 3D object sequences under controlled single or compositional primitive transformations — such as rotation, translation, and scaling — by rendering high-quality scanned meshes from OmniObject3D with Blender.

The dataset was introduced and used in:

• Structure Abstraction and Generalization in a Hippocampal-Entorhinal Inspired World Model — project page
• DiLA: Disentangled Latent Action World Models — project page

Dataset Summary

• Objects: 5,911
• Categories: 216 everyday object categories
• Image resolution: 128 × 128
• Views per object: 72
• Rotation: each object is initialized at 0° and rotated a full 360° around the vertical axis in 5° increments, yielding the 72 rendered views.
• Segmentation masks: one mask is saved per view, enabling synthesis like scaling and translation actions via 2D transformations.

Together, these support controllable transformation sequences: the rendered views give true 3D rotation directly, while the per-view masks let you synthesize additional 2D transformations (e.g., scaling and translation) without any re-rendering. All of these transformations can be applied on their own or composed into combined sequences.

Directory Structure

obj_png_128_72_6000/
├── anise/                       # one folder per category
│   ├── anise_001/               # one folder per object instance
│   │   ├── Scan/
│   │   │   ├── 000.png          # rendered view at 0°
│   │   │   ├── 000_mask.npy     # segmentation mask for view 000
│   │   │   ├── 001.png          # rendered view at 5°
│   │   │   ├── 001_mask.npy
│   │   │   └── ...              # ... up to 071 (355°)
│   │   └── metadata.json        # per-object metadata
│   ├── anise_002/
│   │   └── ...
│   └── ...
└── ...                          # other categories

Data Fields

File	Format	Description
NNN.png	PNG, 128×128	Rendered RGB view; NNN ∈ 000–071, where view index i corresponds to a rotation of 5° × i around the vertical axis.
NNN_mask.npy	NumPy array	Object segmentation mask for the matching view.
metadata.json	JSON	Per-object metadata.

Usage

Download the dataset with the huggingface_hub library:

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="<your-username>/omni-primitive-transforms",
    repo_type="dataset",
)

Load a single view and its mask:

import numpy as np
from PIL import Image

img  = Image.open(f"{local_dir}/obj_png_128_72_6000/anise/anise_001/Scan/000.png")
mask = np.load(f"{local_dir}/obj_png_128_72_6000/anise/anise_001/Scan/000_mask.npy")

Source Data and Rendering

All raw scans are taken from the official OmniObject3D website. We include all raw scans provided on the official website; consequently, the number of categories and objects may differ slightly from those reported in the original OmniObject3D paper. The rendering pipeline is adapted from the implementation released with Objaverse (Deitke et al., 2023).

Code & Reproducibility

The rendering pipeline used to generate this dataset is included in this repository under rendering/; all rendering settings are specified in the script itself. It is adapted from the Objaverse rendering script (Deitke et al., 2023), with the upstream license header and attribution retained.

Setup

Download Blender 3.2.2:

wget https://download.blender.org/release/Blender3.2/blender-3.2.2-linux-x64.tar.xz && \
  tar -xf blender-3.2.2-linux-x64.tar.xz && \
  rm blender-3.2.2-linux-x64.tar.xz

On a headless Linux server, install and start Xorg:

sudo apt-get install xserver-xorg -y && \
  sudo python3 start_x_server.py start

Install the Python dependencies (Python > 3.8 is required). These are the same dependencies as the objaverse-xl rendering scripts.

Reproduce the Dataset

1. Download and unzip the OmniObject3D raw scans.
2. Run the rendering script:

blender-3.2.2-linux-x64/blender --background --python rendering/blender_script.py -- \
  --object_path <path/to/unzipped/OmniObject3D> \
  --output_dir <path/to/output> \
  --num_renders 72 \
  --engine CYCLES \
  --save_image True \
  --save_mask True

License

This dataset is derived from OmniObject3D. Its use is therefore governed by, and must comply with, the original OmniObject3D license and terms. Please review the OmniObject3D terms before downloading or using this data.

Citation

If you use this dataset, please cite the papers that introduced it, as well as OmniObject3D and Objaverse.

@inproceedings{zhang2026structure,
  title = {Structure Abstraction and Generalization in a Hippocampal-Entorhinal Inspired World Model},
  author = {Zhang, Tianqiu and Lyu, Muyang and Liu, Xiao and Wu, Si},
  booktitle = {Forty-third International Conference on Machine Learning},
  year = {2026},
  url = {https://openreview.net/forum?id=AYXgo5FjYz}
}

@inproceedings{zhang2026dila,
  title     = {{DiLA}: Disentangled Latent Action World Models},
  author    = {Zhang, Tianqiu and Lyu, Muyang and Zhang, Yufan and Fang, Fang and Wu, Si},
  booktitle = {Forty-third International Conference on Machine Learning},
  year      = {2026},
  url       = {https://openreview.net/forum?id=AYXgo5FjYz}
}

@inproceedings{wu2023omniobject3d,
  title={Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation},
  author={Wu, Tong and Zhang, Jiarui and Fu, Xiao and Wang, Yuxin and Ren, Jiawei and Pan, Liang and Wu, Wayne and Yang, Lei and Wang, Jiaqi and Qian, Chen and others},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={803--814},
  year={2023}
}

@inproceedings{deitke2023objaverse,
  title = {Objaverse-{{XL}}: A Universe of {{10M}}+ {{3D}} Objects},
  booktitle = {The Thirty-seventh Annual Conference on Neural Information Processing Systems},
  author = {Deitke, Matt and Liu, Ruoshi and Wallingford, Matthew and Ngo, Huong and Michel, Oscar and Kusupati, Aditya and Fan, Alan and Laforte, Christian and Voleti, Vikram and Gadre, Samir Yitzhak and VanderBilt, Eli and Kembhavi, Aniruddha and Vondrick, Carl and Gkioxari, Georgia and Ehsani, Kiana and Schmidt, Ludwig and Farhadi, Ali},
  year = {2023},
}

Acknowledgements

This dataset is built on top of OmniObject3D, and the rendering pipeline is adapted from Objaverse.