Skip to content
SPECIFICTION
Get listed

PartCrafter

github.com

Free plan WindowsLinuxSelf-hosted
Visit github.com
Spec score 8.3 #8 of 24 in AI 3D Model Generators

What it is

PartCrafter is a free, open-source model that generates multiple 3D parts and objects from a single RGB image. It can handle individual objects and complex scenes, producing distinct meshes with meaningful part structure without requiring a pre-segmented image. Its architecture represents parts with disentangled latent tokens and uses hierarchical attention to pass information within and across parts. Local inference scripts can remove backgrounds and optionally apply style transfer to adapt real-world photos toward the rendered-image domain used for training. An optional vision-language feature can suggest the number of parts; Gemini is the default provider, and the provider architecture can be extended. Pretrained object-level and scene-level weights are hosted on Hugging Face. Installation is from the GitHub repository in a Python environment, and inference requires a CUDA-enabled GPU with at least 8 GB of VRAM. The linked Hugging Face demo is paused, though its page says users can ask the authors to restart it. The repository lists an MIT license and provides the software as-is, without warranty.

Spec sheet

Purpose
PartCrafter generates multiple structured 3D parts and objects from a single RGB image in one pass.Read 9 Oct 2026 · wgsxm.github.io
Deployment
The official instructions install the project from its GitHub repository in a Python environment and run inference scripts locally.Read 9 Oct 2026 · github.com
Integrations
The README documents Gemini for part suggestions and style transfer, and says the provider architecture can be extended with other providers.Read 9 Oct 2026 · github.com
Support
The repository invites users to contact the maintainer by email or open an issue with questions or suggestions.Read 9 Oct 2026 · github.com
License
The repository lists an MIT license.Read 9 Oct 2026 · github.com
Architecture
Its method uses disentangled latent tokens for parts and a hierarchical attention mechanism for information flow within and across parts.Read 9 Oct 2026 · wgsxm.github.io
Compositional architecture
The model represents each 3D part with disentangled latent tokens and uses hierarchical attention across parts.Read 9 Oct 2026 · wgsxm.github.io
Demo availability
The linked Hugging Face demo is paused, and its page says users can ask its authors to restart it.Read 9 Oct 2026 · huggingface.co
Hardware requirement
The README specifies a CUDA-enabled GPU with at least 8 GB of VRAM for inference.Read 9 Oct 2026 · github.com
Image preprocessing
The inference scripts support background removal and optional style transfer to adapt real-world photos to the rendered-image domain used for training.Read 9 Oct 2026 · github.com
Image styling
An optional Gemini-powered style-transfer step converts real-world photos toward an Objaverse-style rendering before generation.Read 9 Oct 2026 · github.com
License and warranty
The repository uses the MIT License and provides the software “AS IS,” without warranty.Read 9 Oct 2026 · github.com
Model weights
Pretrained object-level and scene-level model weights are hosted on Hugging Face.Read 9 Oct 2026 · huggingface.co
Object generation
The repository provides an inference script for generating part-level 3D objects from images.Read 9 Oct 2026 · github.com
Part suggestions
An optional VLM feature can suggest the number of parts, using Gemini by default.Read 9 Oct 2026 · github.com
Part-aware output
It generates semantically meaningful, geometrically distinct 3D meshes without requiring pre-segmented input images.Read 9 Oct 2026 · wgsxm.github.io
Scene generation
The repository provides a separate inference script for generating 3D scenes from images.Read 9 Oct 2026 · github.com
Security and compliance
The official pages opened provide no security certification or compliance claim.Read 9 Oct 2026 · github.com
Tested environment
The maintainers say they tested installation on Debian 12 with NVIDIA H20 GPUs and provide community-contributed Windows setup guidance.Read 9 Oct 2026 · github.com
Tested setup
The project says its installation was tested on Debian 12 with NVIDIA H20 GPUs and points Windows users to a community guide and fork.Read 9 Oct 2026 · github.com
Training resources
The project describes training scripts suited to 8 NVIDIA H20 GPUs with 96GB VRAM each.Read 9 Oct 2026 · github.com
Training scale
The training scripts are described as suitable for eight NVIDIA H20 GPUs with 96 GB of VRAM each.Read 9 Oct 2026 · github.com

Plans and prices

PartCrafter (MIT-licensed open-source software) Free MIT license · local installation requires a CUDA-enabled GPU with at least 8GB VRAM Read 9 Oct 2026 · github.com

Every PartCrafter plan and price →

Strong points and limits

Strong points

  • Generates multiple parts from a single RGB image.
  • Supports both individual objects and complex scenes.
  • Does not require pre-segmented input images.
  • Provides pretrained object-level and scene-level weights.
  • MIT-licensed and listed as free.

Limits

  • Local inference requires a CUDA-enabled GPU with at least 8 GB VRAM.
  • The linked Hugging Face demo is paused.
  • Software is provided as-is, without warranty.

Who it suits

PartCrafter suits users who want to generate part-level 3D objects or scenes from images and can run local inference in a Python environment. It requires a CUDA-enabled GPU with at least 8 GB of VRAM.

Verdict

PartCrafter is an image-to-3D model focused on generating distinct, structured parts from one image. Its local GPU requirement and paused demo mean users should be prepared to install and run it themselves.

Where Specifiction places it

Alternatives to PartCrafterAll alternatives →