KaiNinja generates 3D objects as separate parts, not one fused mesh

single source· 1 articles · confidence: medium · first seen 2026-09-13 20:00 UTC

What this means for you

Nothing to act on yet. This is a preprint, the 40% and 16% figures are the authors' own, and no evaluation date is given. If you build part-level pipelines for CAD, rigging or simulation, it is one to watch for a public release.

A preprint on arXiv (2609.15659, 13 September) describes KaiNinja, an extension to TRELLIS.2 that turns a single image into a 3D object made of separate parts rather than one fused mesh. The authors add a dual-volume representation to TRELLIS.2's O-Voxel grid (a voxel holds one sheet of surface, so a single volume cannot show where two parts meet). No segmentation step runs afterwards. Training data includes CAD models and assets written by an LLM-driven agent. The paper reports 40% lower whole-object Chamfer distance — a measure of mesh difference — and 16% higher strict part F-score than other part-generation pipelines; both are self-reported with no evaluation date.

Key facts

  • ·KaiNinja is described in arXiv preprint 2609.15659, dated 13 September 2026, as a part-level extension of the TRELLIS.2 native 3D generator. source
  • ·The method adds a dual-volume representation to TRELLIS.2's O-Voxel grid, which stores one sheet of surface per voxel and therefore cannot represent the interface where two parts touch at any resolution. source
  • ·The pipeline contains no mask or segmenter step. source
  • ·Training data includes CAD models and assets authored by an LLM-driven agent; the authors state it is, to their knowledge, the first 3D generative model trained on agent-authored part data. source
  • ·The authors report it lowers whole-object Chamfer distance by 40% and raises strict part F-score by 16% against part-generation pipelines of other paradigms; no evaluation date is given. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire