# Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers

One‑shot generation of semantically decomposed 3D meshes from a single RGB image

# PartCrafter: One‑Shot Structured 3D Mesh Generation from a Single Image

Imagine uploading a single RGB image and instantly receiving a full 3D model broken down into individual, semantically meaningful parts, ready for editing, printing, or animation. That’s exactly what PartCrafter accomplishes.

## Paper & Repository

**Research Paper**  
_Title:_ “PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers”  
_Authors:_ Yuchen Lin, Chenguo Lin, Panwang Pan, Honglei Yan, Yiqiang Feng, Yadong Mu, Katerina Fragkiadaki  
_Published:_ June 5, 2025, on arXiv

**Official GitHub Repository**  
- Contains full code, with MIT license, and will release model weights + demos by mid-July

**Project Page**  
- Hosted by the authors, summarizing the method and linking to paper and code

## Key Highlights

**One‑Shot Multi-Part Generation**: From a single RGB image, the model generates multiple 3D mesh parts simultaneously, no segmentation needed first.

- **Compositional Latent Space**: Each part corresponds to a distinct set of latent tokens, plus a learned identity embedding.
- **Hierarchical Attention**: A dual attention mechanism maintains detail within parts and coherence across the entire object.
- **Pretrained 3D Mesh Diffusion Transformer (DiT)**: Leverages pretrained DiT components for faster convergence and higher fidelity.

## Performance & Dataset

**Dataset**: Curated ~130,000 objects with part annotations extracted from Objaverse, ShapeNet, ABO with about 300,000 parts.

- **Speed**: Generates full part-aware mesh in ~30–34 seconds on a GPU, outperforming older two-stage pipelines (~18 minutes).
- **Quality**: Achieves lower Chamfer distance and higher F-score (F ≈ 0.7472) than HoloPart and baseline DiT methods

## Use Cases

**3D Printing**: Creates individual STL-ready parts for easy assembly (Tom’s Hardware feature).

- **CAD & Design**: Enables part-level editing and modular asset creation.
- **AR/VR & Games**: Facilitates generation of decomposable assets directly from reference images.
- **Robotics & Simulation**: Provides structural understanding for object manipulation.

## What's Available Now

**Code & Checkpoints**: Public GitHub repo available now. Model weights and demo expected before July 15, 2025.

- **Technical Demo**: Live demonstrations and project page visuals show input-to-mesh workflow.

## Resources

**Project Links & Resources:**

- **Source Code:** [PartCrafter GitHub Repository](https://github.com/wgsxm/PartCrafter)
- **Research Paper:** [Structured 3D Mesh Generation (arXiv)](https://arxiv.org/abs/2506.05573)