Kaininja: Extending Native 3D Generators to the Part Level

Kaininja. Extending Native 3D Generators to the Part Level. By Ruihan Yu and eleven others, from Alaya Lab, the University of Tokyo, and two other institutions. Published on the fourteenth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Native 3D generators turn one image into a single mesh. TRELLIS.2 and its peers deliver high-fidelity non-watertight geometry with materials, but what comes out is one fused object, while downstream work such as editing, rigging and simulation operates on part-level assets. A naive idea is to run a 3D segmentation network on the fused mesh that TRELLIS.2 generates, but such pipelines are slow, and they are bounded by the accuracy of the segmentation. To edit one piece, rig it, or simulate it, the pieces have to exist as separate parts, and a single fused mesh does not provide them.

The authors want a simple way to extend an existing native 3D generator to the part level, and they face a critical problem. The O-Voxel grid stores one sheet of surface per voxel. A single volume therefore cannot represent the interface where two parts touch, and no increase in resolution changes that. Where two parts touch, two surfaces meet, one belonging to each part. A grid that allows one sheet per voxel can hold only one of them, so the boundary between the parts is lost however finely the grid is drawn.

Their solution is a dual-volume representation. KaiNinja is a part-level extension of TRELLIS.2, built on a dual-volume form of its O-Voxel representation, so that the place where two parts meet can be represented. A second volume gives the representation a way to hold what the single volume could not: the interface between parts. And because it is a form of the existing O-Voxel representation, KaiNinja extends TRELLIS.2 rather than replacing it.

KaiNinja keeps the generation speed and quality of TRELLIS.2 while extending it to the part level, and there is no mask and no segmenter in the pipeline. Parts come out of the generator directly. Its training data come from sources of many kinds, including CAD models and assets authored by an LLM-driven agent. To the authors' knowledge, it is the first 3D generative model trained on part data authored by an agent. Mixing CAD models with agent-authored assets widens where part-level training examples can come from.

One result the authors call surprising: whole-object fidelity improves over the same backbone fine-tuned on the same dataset. Moving to parts did not cost quality on the whole object. Against part generation pipelines of different paradigms, KaiNinja lowers whole-object Chamfer distance by forty percent and raises strict part F-score by sixteen percent. Both improvements are measured against pipelines of different paradigms, not against a single kind of competitor.

The obstacle was representational, and so was the fix. One sheet of surface per voxel cannot describe two parts meeting. A second volume can, and with that change the generator produces parts itself, instead of producing one object that a segmenter must cut apart afterwards.