PartCrafter is a structured 3D generative model that reconstructs multiple parts and whole objects from a single RGB image in one shot. Developed by researchers from Peking University and Carnegie Mellon University and accepted to NeurIPS 2025, the project was fully open-sourced under the permissive MIT license and has gathered more than 2,400 GitHub stars from developers and researchers working on image-to-3D generation. ## Part-Aware 3D Generation Most single-image 3D methods output a single fused mesh, which makes the result hard to edit, animate, or reuse part by part. PartCrafter instead generates a 3D object as a set of distinct, semantically meaningful parts from just one input image. That structured output means a chair can come back as a seat, legs, and back rather than one welded shape, giving downstream artists and pipelines components they can manipulate individually. ## Compositional Latent Diffusion The model is built on a compositional latent diffusion transformer. Rather than diffusing a single latent for the entire object, it jointly denoises multiple part latents while modeling the relationships between them, so the parts stay coherent and fit together as a plausible whole. This design lets the network decide how to decompose an object and generate its components in parallel within a single forward pass. ## Single Image to Structured Mesh PartCrafter takes one ordinary RGB image as input and produces a structured 3D mesh, lowering the barrier to 3D asset creation for people without modeling expertise. The repository ships model weights on Hugging Face, an interactive demo Space, and installation guidance, including a community fork with Windows setup instructions. This makes it straightforward to try the approach on custom images without assembling a complex environment from scratch. ## Scene-Level Version Beyond individual objects, the authors released a scene version of PartCrafter trained on the 3D-Front dataset, extending the compositional approach from single objects to multi-object indoor scenes. This broadens the potential use cases from asset generation toward richer scene reconstruction, where several distinct objects need to be recovered and separated at once. ## Considerations PartCrafter is a research release rather than a turnkey production tool: it centers on the paper's official implementation, targets users comfortable with Python and GPU setups, and inherits the accuracy limits common to single-image 3D reconstruction, where occluded regions and fine geometric detail remain challenging. For teams and researchers who specifically need part-separated 3D output from a single image, though, its structured generation approach and permissive MIT license make it a notable and practical open project to build on.