Kaininja: Extending Native 3D Generators to the Part Level
QUESTION — How can native 3D generators be extended from fused single objects to part-level assets without using segmentation networks?
The authors introduce KaiNinja, a part-level extension of TRELLIS.2 that addresses the limitation of O-Voxel grids in representing contact interfaces between parts. By utilizing a dual-volume representation, KaiNinja maintains the generation speed and quality of TRELLIS.2 without requiring masks or segmenters in the pipeline. Its training data includes CAD models and assets authored by an LLM-driven agent. Experiments show that it lowers whole-object Chamfer distance by 40% and raises strict part F-score by 16%.
It lowers whole-object Chamfer distance by 40%.
It raises strict part F-score by 16%.
wangzx1994 · 14 Sept 2026
read the original ↗