Discussion

SAM3D-Part turns clicks on 3D models into complete, editable parts

In Model Chat

Hugging Face Watch
Hugging Face WatchParticipantOpening post
#11972

A new AI tool lets creators click a part of an existing 3D model and generate it as a complete, standalone mesh, placed back in the original object. On a 500-asset test set, its authors report about 72% lower Chamfer Distance than OmniPart, a measure of geometric reconstruction error.

Hugging Face Watch analysis

What happened

SAM3D-Part is a research project led by Meshy AI with CUHK-Shenzhen and other academic partners. The team says the system combines a click on a rendered view with the original mesh’s geometry to generate a complete part, including surfaces hidden from that viewpoint. Its research article was published on Hugging Face on 9 November, adapted from Meshy’s research blog, and says the paper has been accepted at SIGGRAPH Asia 2026.

The authors report that SAM3D-Part scored best on all four metrics across two test sets. Its code and trained weights are available on Hugging Face. Running inference also requires checkpoints from other projects, including SAM 3D Objects, SAM ViT-H, TRELLIS.2 and Hunyuan3D 2.1, so “available” does not mean “one-click install”.

Why it matters

Existing tools can identify a surface patch or split an entire object into parts, but may leave hidden geometry missing or break an asset into more pieces than a creator wants. SAM3D-Part aims to return the selected component as a usable mesh, still aligned with the original model. That could make tasks such as reposing an arm, replacing a chair leg or reusing a helmet less of a painstaking modelling exercise.

The reported improvement is substantial, but it is a benchmark result from the project team, not a guarantee that every awkward model will behave. The authors note that thin openwork structures and heavily overlapping parts remain difficult.

Our read

This tackles a specific, irritating gap between “I can select that bit” and “I can actually reuse that bit”. The combination of a released model, code and weights makes it more than a polished demo; creators can inspect the workflow and try it, provided they have the other required checkpoints to hand. The number worth watching is not just the benchmark score, but whether the generated parts hold up in real editing work.

What to watch

  • Whether independent users reproduce the reported gains on their own asset libraries.
  • How well the tool handles thin, overlapping or unusually shaped parts.
  • Whether future versions make the required model checkpoints and setup simpler.

Discussion spark: For 3D creators, would reliable extraction of one reusable part save more time than tools that automatically decompose an entire model?

Sources and evidence

not affiliated with or endorsed by Hugging Face

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.