Uni-3DAR

Introduction

1. Unified Handling of Multiple 3D Data Types.
   Although we currently focus on microscopic structures such as molecules, proteins, and crystals, the proposed method can be seamlessly applied to macroscopic 3D structures.
2. Support for Diverse Tasks.
   Uni-3DAR naturally supports a wide range of tasks within a single model, especially for both generation and understanding.
3. High Efficiency.
   It uses octree compression-in combination with our proposed 2-level subtree compression-to represent the full 3D space using only hundreds of tokens, compared with tens of thousands in a full-size grid. Our inference benchmarks also show that Uni-3DAR is much faster than diffusion-based models.
4. High Accuracy.
   Building on octree compression, Uni-3DAR further tokenizes fine-grained 3D patches to maintain structural details, achieving substantially better generation quality than previous diffusion-based models.

News

Dependencies

• Uni-Core. For convenience, you can use our prebuilt Docker image:
  docker pull dptechnology/unicore:2407-pytorch2.4.0-cuda12.5-rdma

Reproducing Results on QM9

Download Pretrained Model and Dataset

Inference with Pretrained Model

Train from Scratch

1. Extract the dataset:

1. Run the training script with your desired data path and experiment name:

Citation

@article{lu2025uni3dar,
  author    = {Shuqi Lu and Haowei Lin and Lin Yao and Zhifeng Gao and Xiaohong Ji and Weinan E and Linfeng Zhang and Guolin Ke},
  title     = {Uni-3DAR: Unified 3D Generation and Understanding via Autoregression on Compressed Spatial Tokens},
  journal   = {Arxiv},
  year      = {2025},
}