A joint team from Peking University, ByteDance, and Carnegie Mellon University has released PartCrafter, an open-source generative AI system that turns a single RGB image into multiple structured 3D part meshes in seconds. Detailed in a 5 June 2025 arXiv preprint and a public project demo, the tool uses a compositional latent diffusion transformer to bypass manual segmentation, offering design teams a quicker route from concept photo to fabrication-ready geometry.
Architecture: compositional latent reasoning
Previous efforts either rely on whole-object diffusion models or adopt two-stage “segment-then-reconstruct” workflows such as HoloPart (object level) and MIDI (scene level). These pipelines are sensitive to segmentation errors and incur high computational cost. Other works (Part123, PartGen) reconstruct neural fields rather than explicit meshes, limiting direct use in CAD or simulation. PartCrafter differs by embedding part awareness directly in the diffusion process, eliminating external segmentation steps.
PartCrafter inherits TripoSG weights and encoder/decoder blocks, adding that phrase clarifies reuse. This lets the network learn inter‑part relationships and export up to 16 discrete, non‑overlapping meshes, all aligned to a common coordinate frame. Sixteen parts comfortably covers most household objects in Objaverse while keeping token budgets tractable.
The team fine‑tuned on 50,000 parts‑annotated objects drawn from Objaverse, ShapeNet, and the ABO dataset. On the 3D part‑level object generation benchmark, PartCrafter achieved an L2 Chamfer Distance of 0.1726 on the Objaverse split, improving on HoloPart’s 0.1916, and reduced generation time from 18 minutes to 34 seconds on a single H20 GPU, roughly about a 32 × speed‑up. All generated vertices share a global canonical coordinate frame, allowing direct recombination or editing.
Industry context
PartCrafter arrives amid a surge of AI tools aimed at additive manufacturing. In February, Tencent rolled out its Hunyuan3D 2.0 text‑to‑mesh workflow for full‑colour models, and March saw Tripo AI debut an API that converts prompts directly into editable CAD features, both underscoring industry demand for faster content pipelines.
Elsewhere, Autodesk Research is pursuing a parallel goal with Project Bernini, an experimental generative-AI model that turns text, sketches, or multi-view images into full 3D meshes in seconds.
Future work & roadmap
The authors plan to scale training to millions of annotated parts and embed physical priors so generated assemblies respect real‑world tolerances. Code and data will be released after peer review. Potential use-cases span industrial CAD automation, AR/VR asset pipelines, and physics-based simulation.
Register for our upcoming event Additive Manufacturing Advantage: Aerospace, Space and Defense.
What 3D printing trends should you watch out for in 2025?
How is the future of 3D printing shaping up?
To stay up to date with the latest 3D printing news, don’t forget to subscribe to the 3D Printing Industry newsletter or follow us on Twitter, or like our page on Facebook.
While you’re here, why not subscribe to our Youtube channel? Featuring discussion, debriefs, video shorts, and webinar replays.
Featured image shows generated 3D meshes from RGB images. Image via Lin et al./PartCrafter project page.
