wgsxm

PartCrafter: Generate Structured 3D Meshes from Images

PartCrafter generates part-separated 3D objects and scenes from a single RGB image using compositional latent diffusion. It suits researchers and developers exploring image-to-3D generation who have access to a CUDA-enabled GPU.

View on GitHub Compare
Stars
2.5k
Forks
165
Stars, last 30 days
+20

Our take

Health score

23/100

How it is scored
Activity
19
Community
30
Issues
21
Pull requests
20

+20 stars in the last 30 days

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

2,512

Forks

166

Change since Nov 15, 2026: 0 stars·0 forks

Overview

PartCrafter is a research implementation for turning one RGB image into a structured 3D mesh. Rather than generating an object as a single undivided shape, it generates multiple parts together, and also provides a scene-generation variant.

It is a fit for researchers and developers working on image-to-3D generation, 3D reconstruction, or compositional 3D content. The repository includes inference scripts and pretrained object- and scene-level checkpoints, as well as training code for users who want to investigate or adapt the approach.

Key Features

  • Generates part-level 3D objects from a single image, with a user-specified number of parts.
  • Includes a separate model and inference script for generating 3D scenes.
  • Provides pretrained checkpoints for object and scene generation.
  • Supports optional background removal for custom input images.
  • Offers optional VLM-based part-count suggestions and image style transfer, using Gemini by default.
  • Includes training scripts and dataset preprocessing utilities.
  • Saves generated results locally and can render them as part of inference.

Use Cases

  • Researchers can reproduce or extend work on compositional 3D generation using the released inference and training code.
  • 3D content creators can explore turning a reference image into a mesh with distinct parts, when a CUDA GPU is available.
  • Developers can test image-to-3D generation in object-level or scene-level workflows using the provided checkpoints.
  • Teams building 3D data pipelines can use the preprocessing utilities as a starting point for preparing GLB assets.

Getting Started

The README recommends Python 3.11 and PyTorch 2.5.1 with CUDA 12.4. After installing those prerequisites, clone the repository and install its dependencies:

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

For an example object-generation run, see the inference command and further setup details in the README. The project page provides additional project information.

Alternatives

  • LAM: LAM reconstructs and animates a 3D Gaussian head avatar from a single image, rather than generating part-separated general objects and scenes.
ProjectLanguageLicenseStarsStatus
PartCrafterPythonMIT2.5kNot checked yet
LAMPythonApache-2.01.1kQuiet

Considerations

  • Inference requires a CUDA-enabled GPU with at least 8 GB of VRAM. Reducing the number of parts or tokens can lower memory use.
  • The documented environment uses Python 3.11 and PyTorch with CUDA 12.4. Installation was tested on Debian 12 with NVIDIA H20 GPUs; Windows setup guidance is linked from the README.
  • Training is substantially more demanding than inference: the authors describe configurations for 8 H20 GPUs with 96 GB VRAM each.
  • The project notes that models were trained on rendered images. Real-world photos may benefit from the optional style-transfer step, which uses a VLM provider and requires an API key.
  • Preprocessed dataset release is listed as unfinished in the README, so preparing training data may require additional work.

Found this useful?

Share it with someone who would like PartCrafter.

Comparisons