---
license: mit
pipeline_tag: image-to-3d
tags:
  - 3d-generation
  - part-aware
  - trellis
---

# KaiNinja

Weights of KaiNinja, a part-aware image-to-3D generator built on TRELLIS.2.

| File | Stage | Model | Weights |
|---|---|---|---|
| `stage1_layout_flow.safetensors` | Stage 1, dual-volume sparse-structure flow | `TriSparseStructureFlowModel` | step 75k |
| `stage2_slat_flow.safetensors` | Stage 2, dual-stream structured-latent flow | `TriSLatSeparateFlowModel` | EMA (0.9999), step 100k |

The matching `*.json` files hold the model constructor arguments. The shape decoder is the original
TRELLIS.2 one and is not included here.

## Inference settings

- Stage 1: Euler flow sampler, 50 steps, classifier-free guidance 3.0; the two decoded volumes go
  through overlap deduplication before Stage 2.
- Stage 2: Euler flow sampler, 50 steps, classifier-free guidance 3.0.
- Each volume is decoded with the TRELLIS.2 shape decoder; parts are the connected components within
  each volume.

## License

MIT, as the base model.