Native 3D generators such as TRELLIS.2 can turn a single image into a high-fidelity mesh with materials, but the result is one fused object. Downstream tasks like editing, rigging and simulation, however, need assets split into parts. Running a 3D segmentation network on the fused mesh is one option, yet such pipelines are slow and limited by segmentation accuracy.
Researchers introduce KaiNinja, a part-level extension of TRELLIS.2 built on a dual-volume form of its O-Voxel representation. The core obstacle is that an O-Voxel grid stores only one sheet of surface per voxel, so a single volume cannot represent the interface where two parts touch at any resolution. The dual-volume representation is designed to resolve this interface conflict.
According to the paper, KaiNinja preserves the generation speed and quality of TRELLIS.2 while adding part-level output, with no mask or segmenter in the pipeline. Its training data come from many sources, including CAD models and assets authored by an LLM-driven agent; the authors state this is the first 3D generative model trained on agent-authored part data.
Notably, the team reports that whole-object fidelity also improves over the same backbone fine-tuned on the same dataset. Compared with part generation pipelines of different paradigms, KaiNinja lowers whole-object Chamfer distance by 40% and raises strict part F-score by 16%.