Introduction

Native image-to-3D models finally ship sharp, PBR-ready meshes. Then production reality hits: editing, rigging, and simulation want parts, not one fused shell. Run a segmenter after the fact and you pay latency plus segmentation error. Bake parts into the generator and you fight a harder representation problem — shared contact surfaces do not fit cleanly in a single sheet-per-voxel volume.

KaiNinja (arXiv 2609.15659, on Hugging Face Daily Papers September 15, 2026) extends TRELLIS.2 to the part level with a dual-volume O-Voxel representation. No mask. No post-hoc segmenter. Part fidelity jumps, and whole-object quality improves on the same backbone. If your pipeline lives on image-to-3D for game props, this is the control surface that turns pretty meshes into shippable assets.

The Magic: Dual Volumes Beat Fused Shells

TRELLIS.2 and peers turn one image into a high-fidelity non-watertight mesh with materials. Great for lookdev. Weak for production: a fused object cannot expose independent blades, panels, joints, or lids without expensive cleanup.

The blocker is geometric, not cosmetic. O-Voxel stores one surface sheet per voxel. Where two parts touch, a single volume cannot represent that interface at any resolution. Segment-then-rebuild pipelines paper over the issue — slowly, and only as well as the segmenter.

KaiNinja's move:

  1. Pack object parts into two interleaved volumes (contact graph contracted until bipartite).
  2. Run a two-stream generation path over TRELLIS.2's structured latents.
  3. Decode part-level assets while keeping TRELLIS.2 speed and quality.
  4. Train on mixed sources — CAD plus LLM-agent-authored part data (claimed first of its kind for a 3D generative model).

Headline numbers against competing part-generation paradigms: 40% lower whole-object Chamfer distance, 16% higher strict part F-score. Surprise bonus: whole-object fidelity beats the same backbone fine-tuned on the same dataset without the dual-volume split. Parts are not a quality tax here — they are a better inductive bias.

Code and checkpoints ship under the PartTrellis release packaging (TRELLIS.2 O-Voxel dual-volume stack), with inference scripts such as end-to-end dual decoding and part evaluation helpers.

Pro Tip

Design your contact graph for bipartite packing early. If two parts must share a thick welded interface, treat the weld as its own part or accept a soft boundary. Dual volumes solve co-located surfaces; they do not invent articulation constraints. Name parts for downstream rigging the moment you export — anonymous mesh islands still waste an animator day.

Wire Part-Level Generation Into Your Asset Loop

Use this checklist when you bolt KaiNinja-style part generation onto a TRELLIS.2 image-to-3D path:

  1. Start from a single reference image of the prop or character piece.
  2. Run the dual-volume part generator (not a segmenter on a fused mesh).
  3. Export named part meshes plus the assembled whole for QA.
  4. Verify contact interfaces: no z-fighting, no missing cavity walls, no glued hinges.
  5. Feed parts into your usual DCC pipeline — UV, LODs, colliders, sockets.
  6. Benchmark against your segmenter baseline on Chamfer / part F-score before swapping production.

# Orientation sketch (adapt to the PartTrellis / KaiNinja release layout)
# 1) Install TRELLIS.2 + O-Voxel stack
# 2) Load dual-volume part weights
# 3) Infer parts from one image, then assemble for QA

python scripts/infer_e2e_dual.py \
  --image ./refs/prop.png \
  --out ./exports/prop_parts \
  --save-parts --save-whole

# Production gate before you trust a batch
check:
  - part_count matches expected BOM
  - contact faces resolve without double shells
  - whole-object CD vs fused TRELLIS.2 baseline
  - articulating parts expose clean pivots for rigging

Paper: arXiv 2609.15659 · HF Daily: huggingface.co/papers/2609.15659 · Weights packaging: huggingface.co/AuroraRyan/PartTrellis · Backbone: TRELLIS.2

Conclusion

KaiNinja turns TRELLIS.2 from a pretty fused mesh machine into a part-native generator that game pipelines can actually consume. Dual volumes fix the contact interface that single O-Voxel sheets cannot express, skip the segmenter tax, and still lift whole-object fidelity. If your image-to-3D props still arrive as one welded lump, you are shipping lookdev and calling it production. Generate the parts. Keep the speed.

—Titus