KaiNinja: Extending Native 3D Generators to the Part Level
Ruihan Yu, Lian Fu, Muyao Niu, Zheng-hui Huang, Yu-Ju Tsai, Sho Kuno, Fengbo Lan, Yonghao Yu, et al.
cs.GR · cs.AI · cs.CV
Quick verdict
"KaiNinja extends TRELLIS.2 to part-level 3D generation via dual-volume packing; results are strong but the part-level edge shrinks or reverses on the external benchmark."
Summary
KaiNinja extends the native 3D generator TRELLIS.2 to part-level output via a dual-volume O-Voxel packing built on contact-graph two-coloring, so touching parts land in different volumes; from one image it outputs parts as separate meshes with no mask or segmenter in roughly 24 seconds per object on an H100. The layout flow receives per-stream weights with zero-initialized cross-volume attention, while the refinement flow keeps the pretrained backbone and separates streams by attention scope. Training uses 19,132 part-annotated objects from four sources, including LLM-agent-authored Articraft-10K. On a 986-object held-out split, it ranks best on all seven metrics, cutting whole-object Chamfer distance by 40% and raising strict part F-score by 0.095 against the strongest cascade, and beats the same backbone fine-tuned on the same corpus on whole-object fidelity. On Sketchfab/GitHub assets the whole-object lead holds, but on external HY3D-Bench the part-level ordering reverses, with Hunyuan3D-2.1+X-Part winning every part metric. Limitations include undersegmentation when non-bipartite contact graphs force parts to merge, and three failure modes, including single-volume collapse on about 0.9% of test objects.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A dual-volume packing of the O-Voxel representation lets TRELLIS.2 generate part-separated assets in one pass with no mask or segmenter, because touching parts never share a volume.
- On a 986-object held-out test set, KaiNinja ranks best on all seven metrics of Table 1, including 40% lower whole-object Chamfer distance and +0.095 strict part F-score against the strongest cascade (Hunyuan3D-2.1+X-Part).
- Dual-volume packing improves whole-object fidelity beyond the same backbone fine-tuned on the same corpus (0.919 vs 0.830 F1@0.05W; 0.7% vs 1.8% failure rate).
- Generation runs in roughly 24 seconds per object on one H100, an order of magnitude faster than the generate-then-segment cascade.
- To the authors' knowledge it is the first 3D generative model trained on agent-authored 3D assets (Articraft-10K), whose part labels are exact by construction.
Strengths
Clear representation-level problem statement
Section 3.1 argues a single O-Voxel volume cannot represent a part interface because each voxel stores one surface sheet, so contact faces collapse; dual-volume packing is presented as the minimal remedy, with Appendix B arguing why the PartPacker planarity/four-color suggestion fails (Tietze/Besicovitch).
Complete and honest results reporting
Section 4 spans the main comparison (Table 1), per-source breakdown (Table 3), whole-object references (Table 2), two external sets (Tables 4-5), and full ablations (Tables 6-7), including unfavorable results such as the part-metric reversal on HY3D-Bench and 9/1000 single-volume collapses.
Well-controlled headline whole-object comparison
Novel, exactly-labeled training source
Weaknesses
In-distribution advantage on the headline part-level comparison
Custom evaluation protocol with method-favoring choices
The shipped model is not the best-scoring configuration
Better arms left unevaluated
Compute and stochasticity never quantified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
KaiNinja extends TRELLIS.2, a native (sparse-voxel, flow-based) image-to-3D generator, to part-level output. The enabling observation (Section 3.1) is representational: an O-Voxel voxel stores at most one surface sheet, so the two nearly-parallel faces where two parts touch collapse into one voxel at any resolution. The remedy is a dual-volume packing - two-coloring the part-contact graph, with greedy odd-cycle contraction for non-bipartite cases - so touching parts land in different volumes and parts can later be read off as connected components, with no mask or segmenter on the critical path. Central claims: (1) state-of-the-art part-level image-to-3D on the paper's 986-object held-out test set (all seven Table 1 metrics, CD_W 0.0186 vs 0.0314 for the strongest cascade, F1@0.05P 0.692 vs 0.597); (2) improved whole-object fidelity over the same backbone fine-tuned on the same corpus (Table 2); (3) ~24 s/object on an H100; (4) the first 3D generative model trained on agent-authored part data (Articraft-10K), appropriately hedged.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is KaiNinja?
KaiNinja is a research model that extends the 3D generator TRELLIS.2 so that, from a single image, it outputs a 3D asset whose individual parts are separate meshes. It uses a dual-volume voxel packing that keeps touching parts in different volumes, so no segmentation network or mask is needed.
How fast is KaiNinja?
The paper reports roughly 24 seconds per object on one NVIDIA H100 at 512 resolution, much faster than pipelines that generate a whole mesh first and then run a segmentation network, which the authors measure at roughly ten times the cost.
Does KaiNinja beat X-Part and PartPacker?
On the authors' held-out test set of 986 objects it ranks best on all reported quality metrics. On the external HY3D-Bench it still wins on whole-object geometry, but on part-level metrics the Hunyuan3D-2.1+X-Part cascade scores higher, because KaiNinja cuts objects into more pieces than that benchmark's labels reward.
Is code available?
The paper lists a project page (alaya-lab.github.io/KaiNinja) and a GitHub repository (github.com/AlayaLab/KaiNinja); whether the repository currently contains weights and full training code should be verified on the linked page.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-10 cannot be assessed visually from the extracted text; qualitative claims about part quality rest on the authors' descriptions.
- KaiNinja's own wall-clock cost appears only as a prose figure (about 24 s/object) and not in any table.
- The code URL is stated in the paper but its contents were not verified.
- The abstract's '16%' strict part F-score figure corresponds to +0.095 absolute points in Table 1; the derivation of the percentage is not spelled out.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.