KaiNinja Paper Extends TRELLIS.2 3D Generator to Part Level

KaiNinja, a part-level extension of TRELLIS.2, lowers whole-object Chamfer distance by 40% and raises strict part F-score by 16% against competing pipelines, according to a new preprint from Ruihan Yu, Ming-Hsuan Yang and collaborators.

Native 3D generators like TRELLIS.2 turn an image into one fused mesh, which the authors note is a mismatch with downstream work: "editing, rigging and simulation operates on part-level assets." A common workaround is to run a segmentation network over the finished mesh, but the abstract calls those pipelines "slow and bounded by the accuracy of the segmentation." The blocker to a native fix is representational: "the O-Voxel grid stores one sheet of surface per voxel, so a single volume cannot represent the interface where two parts touch, at any resolution." KaiNinja's answer is a dual-volume form of the same grid, letting the model emit distinct part meshes "with no mask or segmenter in the pipeline."

Training data comes from CAD models and assets "authored by an LLM-driven agent"; the authors describe KaiNinja as "the first 3D generative model trained on agent-authored part data." A side finding the paper labels "surprisingly": whole-object fidelity improves over the same backbone fine-tuned on the same dataset without the part-level extension. The abstract does not name the baselines behind the 40% and 16% figures or publish per-category numbers, and lands amid a busy stretch on our computer-vision tracker.