Image-to-3D generators have gotten genuinely good. Point TRELLIS.2 or a peer at a photo and you get back detailed geometry with materials, in seconds. Then you open it in Blender and hit the real problem: it’s one fused object.
A chair whose legs aren’t separate from its seat can’t be rigged. A robot whose arm is welded to its body can’t be posed. A machine whose cover isn’t its own mesh can’t be opened. Rigging, editing, animation and simulation all operate on parts, and generated meshes don’t have any.
KaiNinja (arXiv, submitted Sept 14, revised Sept 15) is a direct fix, with code and a project page released.
Why segmentation afterwards doesn’t work
The obvious approach is to generate the fused mesh and then run a 3D segmentation network over it to cut it up. The paper’s objection is practical: those pipelines are slow, and the quality of your parts is capped by the accuracy of the segmenter. You’ve added a second failure point downstream of the first.
The actual obstacle
The interesting part of this paper is the problem diagnosis. TRELLIS.2 represents geometry in an O-Voxel grid, and each voxel stores one sheet of surface. That’s efficient, and it’s why the output is high quality.
It also makes part-level output structurally impossible. Where two parts touch — the drawer against the cabinet, the arm against the shoulder — you need two surfaces in the same place: the outside of one part and the outside of the other. One sheet per voxel cannot express that, and no amount of extra resolution fixes it, because it’s a representational limit rather than a precision limit.
That’s a clean example of something worth recognizing generally: when a generative system fails at something, it’s often because its representation cannot hold the answer, not because it needs more training.
The fix
KaiNinja introduces a dual-volume representation: two volumes, so the interface between touching parts can be represented from both sides. It keeps TRELLIS.2’s generation speed and quality and adds part-level output, with no mask and no segmenter in the pipeline.
The training data is notable too. It comes from varied sources including CAD models and, per the authors, assets authored by an LLM-driven agent — one of the first 3D generative systems to be trained partly on agent-authored geometry.
Why it matters outside research
If it holds up, this shortens the gap between “generated a thing” and “using the thing.” Concept-to-asset workflows currently stall at the cleanup step: retopology, splitting, rigging. Parts that arrive already separated skip the most tedious part of that, particularly for props, hard-surface objects and anything mechanical.
Worth watching whether it appears in Blender or ComfyUI wrappers, which is how this class of research reaches working artists.