Text- and image-to-3D has got good enough that the bottleneck moved. The meshes look right. The problem is what they are: a single fused object, one continuous surface with no internal structure. You cannot rig it, you cannot swap a part, you cannot animate a lid opening, and editing means sculpting rather than selecting.
KaiNinja, posted to arXiv on September 14, 2026 (revised the 15th) by a large team including Ruihan Yu, Lian Fu, Muyao Niu, Erwin Wu, Ming-Hsuan Yang, Kaipeng Zhang and Zhixiang Wang, makes the generator produce parts in the first place.
The obstacle
Native 3D generators like TRELLIS.2 represent objects in an O-Voxel grid — a sparse occupancy volume the model fills in. The limitation is structural: a single O-Voxel grid cannot represent the interface between two adjacent parts. Where a handle meets a mug, there is one occupied region, and nothing in the representation distinguishes “handle surface” from “mug surface” from “the seam between them.”
The common workaround is to generate the fused mesh and then run a segmentation network over it to guess where the parts are. That’s a second model, a second failure mode, and it’s guessing at structure that the generator destroyed.
The fix
KaiNinja is described as “a part-level extension of TRELLIS.2 built on a dual-volume form of its O-Voxel representation.” Rather than one volume, two — which gives the representation somewhere to encode the interface between adjacent parts, the thing a single grid cannot express.
The consequences the authors emphasise:
- No segmentation network. Part decomposition falls out of the representation instead of being recovered after the fact.
- Generation speed is maintained. This is not a slower, more careful pipeline; it’s the same generator with a richer volume.
Numbers
Against competing part-level pipelines, KaiNinja reports:
- ~40% reduction in whole-object Chamfer distance — the generated surface is closer to ground truth overall, not just better segmented.
- ~16% improvement in strict part F-score — the parts themselves are more correct.
The first of those is the more interesting result. You would expect a part-aware method to trade some whole-object fidelity for structure; instead the whole object got substantially better. That suggests the part interfaces were carrying information the single-volume representation was throwing away — that modelling the seams helps you model the shape.
Why a creative audience should care
Generated 3D has been stuck in a specific role: background props, greyboxing, things nobody will need to touch. Everything a 3D artist actually does downstream — rigging, UV work, material assignment per component, animating mechanisms — assumes the model has parts.
A generator that emits a part-level decomposition natively, fast, moves generated assets from “reference geometry” toward “starting point.” That is a different tool. Not a finished asset, but the difference between inheriting a block of marble and inheriting a rough-cut kit.
The authors have published a project page and code under CC-BY 4.0, which means this is checkable rather than merely claimed — worth doing before believing the numbers, as with any preprint.