README.md

metadata

license: mit

Uni-3DAR

[Paper]

Introduction

Schematic illustration of the Uni-3DAR framework

Uni-3DAR is an autoregressive model that unifies various 3D tasks. In particular, it offers the following improvements:

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.

Usage

Please visit our GitHub Repo (https://github.com/dptech-corp/Uni-3DAR) for detailed instructions.