---
license: apache-2.0
tags:
  - 3d
  - mesh-generation
  - auto-rigging
  - onnx
---

# Syntverk Mesh Models (ONNX)

Pre-exported ONNX models for 3D mesh generation and auto-rigging in
[Syntverk](https://syntverk.com), a native agentic coding tool.

## Models

### Hunyuan3D-2mini (Shape Generation)
Image → 3D mesh via flow-matching diffusion.

| File | Size | Description |
|------|------|-------------|
| `dinov2_encoder.onnx` + weights | ~3.5 GB | DINOv2-giant image encoder (fp16) |
| `shape_dit.onnx` | ~3 MB | Diffusion transformer (fp16) |
| `vae_transformer.onnx` | ~385 MB | VAE latent decoder (fp16) |
| `vae_cross_attn_decoder.onnx` | ~25 MB | VAE geometry decoder (fp16) |

Original model: [Tencent/Hunyuan3D-2mini](https://huggingface.co/Tencent/Hunyuan3D-2mini) (Apache 2.0)

### Puppeteer (Auto-Rigging)
Mesh → skeleton + skinning weights.

| File | Size | Description |
|------|------|-------------|
| `skeleton_encoder.onnx` | ~245 MB | Michelangelo point cloud encoder (fp32) |
| `skeleton_gpt.onnx` | ~580 MB | SkeletonGPT OPT-350M transformer (fp16) |
| `skinning_net.onnx` | varies | SkinningNet with PartField + TAJA |
| `weights/*.npy` | ~11 MB | Embedding weights for token generation |
| `puppeteer_config.json` | <1 KB | Model configuration |

Original model: [Seed3D/Puppeteer](https://huggingface.co/Seed3D/Puppeteer) (Apache 2.0)

## Usage

These models are automatically downloaded by Syntverk on first use.
Place them in `~/.local/share/syntverk/models/mesh/`.

```
~/.local/share/syntverk/models/mesh/
├── hunyuan3d-mini/
│   ├── dinov2_encoder.onnx
│   ├── ... (weight files)
│   ├── shape_dit.onnx
│   ├── vae_transformer.onnx
│   └── vae_cross_attn_decoder.onnx
└── puppeteer/
    ├── skeleton_encoder.onnx
    ├── skeleton_gpt.onnx
    ├── skinning_net.onnx
    └── weights/
        └── *.npy
```

## License

The ONNX exports are derived from models under Apache 2.0:
- Hunyuan3D-2mini: Copyright Tencent
- Puppeteer: Copyright ByteDance
- Michelangelo: Copyright original authors
- PartField: Copyright original authors

This repository packages those weights in ONNX format for use with ONNX Runtime.