PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers

PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers

Summary

PartCrafter is an innovative structured 3D generative model that enables the creation of complex 3D meshes and scenes from a single RGB image. Accepted at NeurIPS 2025, this project leverages compositional latent diffusion transformers to jointly generate multiple parts and objects in one shot. It offers powerful capabilities for both 3D object and scene generation, making it a valuable tool for researchers and developers in the field.

Repository Info

Updated on January 14, 2026
View on GitHub

Tags

3d generation3d meshlatent diffusiontransformersimage to 3dpythongenerative AIneurips

Click on any tag to explore related repositories

Introduction

PartCrafter is a groundbreaking open-source project that introduces a structured 3D generative model capable of creating intricate 3D meshes and entire scenes from a single 2D RGB image. This research, accepted at NeurIPS 2025, utilizes Compositional Latent Diffusion Transformers to achieve its impressive results. Unlike traditional methods, PartCrafter can jointly generate multiple parts and objects in a single pass, offering a more holistic and efficient approach to 3D content creation. It provides robust solutions for both part-level 3D object generation and comprehensive 3D scene generation.

Installation

To get PartCrafter up and running, follow these steps. The project recommends torch-2.5.1+cu124 and python-3.11, though other versions might also work. A CUDA-enabled GPU with at least 8GB VRAM is required.

First, create a conda environment (optional):

conda create -n partcrafter python=3.11.13
conda activate partcrafter
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu124

Next, clone the repository and install dependencies:

git clone https://github.com/wgsxm/PartCrafter.git
cd PartCrafter
bash settings/setup.sh

If you do not have root access and are using a conda environment, you can install required graphics libraries with:

conda install -c conda-forge libegl libglu pyopengl

For Windows users, specific installation guides are available in this pull request and this fork.

Examples

PartCrafter offers quick start scripts for both 3D object and scene generation. The necessary model weights are automatically downloaded upon first use.

3D Part-Level Object Generation

Generate a 3D part-level object from an image:

python scripts/inference_partcrafter.py \
  --image_path assets/images/np3_2f6ab901c5a84ed6bbdf85a67b22a2ee.png \
  --num_parts 3 --tag robot --render

Specify --rmbg if you use custom images. This will remove the background of the input image and resize it appropriately.

3D Scene Generation

Generate a 3D scene from an image:

python scripts/inference_partcrafter_scene.py \
  --image_path assets/images_scene/np6_0192a842-531c-419a-923e-28db4add8656_DiningRoom-31158.png \
  --num_parts 6 --tag dining_room --render

Why Use PartCrafter?

PartCrafter stands out for several compelling reasons:

  • Cutting-edge Research: Recognized and accepted at NeurIPS 2025, showcasing its significant contributions to the field of generative AI.
  • Structured 3D Generation: It uniquely generates 3D objects and scenes with distinct, compositionally structured parts, offering more control and realism.
  • Image-to-3D Capability: Seamlessly transforms 2D RGB images into complex 3D models and environments, streamlining the content creation pipeline.
  • Advanced AI Architecture: Leverages Compositional Latent Diffusion Transformers, a sophisticated approach for high-quality and coherent 3D outputs.
  • Versatility: Supports both detailed part-level object generation and expansive 3D scene construction, catering to diverse application needs.
  • Open-Source and Accessible: Fully open-sourced with pre-trained models available on HuggingFace, including an interactive demo for easy experimentation.

Links