KaiNinja: Extending Native 3D Generators to the Part Level

KaiNinja is a part-level extension of native 3D generation, aimed at one of the biggest practical gaps in current image-to-asset workflows: most generators still output a single fused mesh. That works for previewing a model, but it becomes a liability once the asset needs to be edited, rigged, broken apart, or fed into simulation and gameplay systems.

The core technical change is a dual-volume representation built on top of TRELLIS.2’s O-Voxel approach. The authors argue that a single volume cannot properly represent the interface where two parts meet, so they split the representation to preserve part boundaries without adding a separate segmentation stage. That matters because a post-process segmenter adds latency and introduces another failure point.

KaiNinja is trained on a mix of CAD models and assets authored by an LLM-driven agent, which the team says makes it the first 3D generative model trained on agent-authored part data. In practice, that suggests a pipeline that can learn part structure from more than just clean CAD sources, which is relevant for teams trying to bridge authored assets and generative workflows.

The reported results are notable for production-minded users: compared with part-generation pipelines using different approaches, KaiNinja lowers whole-object Chamfer distance by 40% and improves strict part F-score by 16%. Just as interesting, whole-object fidelity reportedly improves over the same backbone fine-tuned on the same dataset, hinting that part awareness may help rather than hurt overall mesh quality.