Kaininja: Extending Native 3D Generators to the Part Level
KaiNinja changes TRELLIS.2 so it can generate separate 3D parts directly, without adding a segmentation step.
The paper says the issue is structural: TRELLIS.2-style O-Voxel grids cannot represent touching part interfaces inside one volume. KaiNinja uses a dual-volume representation to keep those interfaces separable while preserving the native generator’s speed and quality. Its training data includes CAD sources and part assets made by an LLM-driven agent. The authors report a 40% lower whole-object Chamfer distance and a 16% higher strict part F-score against other part-generation pipelines. HF Daily Papers' note
The paper says the issue is structural: TRELLIS.2-style O-Voxel grids cannot represent touching part interfaces inside one volume. KaiNinja uses a dual-volume representation to keep those interfaces separable while preserving the native generator’s speed and quality. Its training data includes CAD sources and part assets made by an LLM-driven agent. The authors report a 40% lower whole-object Chamfer distance and a 16% higher strict part F-score against other part-generation pipelines. HF Daily Papers' note
score 5