PartCrafter object and scene models

PartCrafter gives an artist separate meshes to work on from a single reference image. The object checkpoint generates parts of one object; the scene checkpoint generates multiple objects in a scene. Furniture, equipment and assemblies are sensible candidates when splitting a single mesh would be tedious. Its value depends on how much cleanup those generated pieces save. They still need meaningful names, pivots, intersection checks and material work before they become a usable game asset. A drawer exported as its own file still needs someone to make it open properly. The supplied inference scripts export individual GLB meshes alongside a combined model, giving that cleanup a practical starting point.

Commercial use & license

The project license and both official checkpoint cards say MIT. Preserve the applicable notices for TripoSG components and DINOv2's Apache-2.0 terms. The object example also loads BRIA RMBG-1.4: free use is non-commercial, and commercial use requires a separate BRIA agreement. Optional Gemini preprocessing has separate service terms.

Adult content rules

The project license does not establish rights to reference images, likenesses or third-party assets. The object example loads BRIA RMBG-1.4, whose free use is non-commercial; its linked 1.4 agreement returned 404, so confirm the applicable terms before commercial or adult use. Optional Gemini preprocessing falls under Google's prohibition on sexually explicit content, subject to its discretionary exceptions. Other hosted deployments have separate policies. No blanket adult-content permission is claimed.

Read the policy ↗

What you need

Official inference calls for a CUDA-enabled GPU with at least 8 GB VRAM. Part count and tokens per part affect memory; the examples default to 1,024 tokens per object part and 2,048 per scene object. The documented setup uses Python 3.11 and PyTorch 2.5.1 with CUDA 12.4.

Ranks are recommendations for the task described, based on reviewed documentation. Model ranks cover non-explicit development work. We have not run practical benchmarks. Check content policies and commercial terms separately. How the ranks work →