sf3d-webgpu-weights / README.md
lyonsno's picture
Add SF3D WebGPU weights, license, notice, and model card
53cfd89 verified
|
Raw History Blame Contribute Delete
2.71 kB
---
license: other
license_name: stability-ai-community
license_link: LICENSE.md
base_model: stabilityai/stable-fast-3d
tags:
- 3d-generation
- image-to-3d
- webgpu
- stable-fast-3d
library_name: webgpu
---
# SF3D WebGPU Weights
WebGPU-ready weights for [**SF3D WebGPU**](https://github.com/lyonsno/sf3d-webgpu) —
a browser port of Stability AI's [Stable Fast 3D](https://huggingface.co/stabilityai/stable-fast-3d)
that runs single-image → textured 3D mesh generation entirely in WebGPU compute
shaders. No server, no Python, no ONNX at inference time.
This repository hosts the converted model weights so you can skip the PyTorch
conversion step: **clone the app, download `weights.bin`, run.**
> **Powered by Stability AI.**
## Files
| File | Description |
|------|-------------|
| `weights.bin` | Flat fp16 binary (~2.1 GB) for direct WebGPU loading |
| `LICENSE.md` | Stability AI Community License (governs these weights) |
| `NOTICE` | Attribution + derivative-work modification notice |
```
SHA-256(weights.bin) = 0e5c23c8c502492c0b4432006ac30f2d11b2a4c102a3bf9f21c2fcf094ddbffd
```
## Usage
```bash
git clone https://github.com/lyonsno/sf3d-webgpu
cd sf3d-webgpu
npm install
# Download these weights into public/
hf download BasinShapers/sf3d-webgpu-weights weights.bin --local-dir public/
npm run dev # open http://localhost:5177/
```
On an M4 Max the demo image generates a textured GLB in ~33s. See the
[application repo](https://github.com/lyonsno/sf3d-webgpu) for the full pipeline,
architecture, and PyTorch-parity receipts (the demo output is byte-for-byte
reproducible).
## What these weights are
This is a **derivative work** of Stability AI's Stable Fast 3D: a format
conversion of the original `model.safetensors` into a single flat binary laid
out for WebGPU compute shaders (fp16, transposed linear weights, WebGPU-relevant
tensors only). The model architecture and learned values are unchanged. Full
details are in [`NOTICE`](NOTICE); the conversion tool is
[`tools/convert_weights.py`](https://github.com/lyonsno/sf3d-webgpu/blob/main/tools/convert_weights.py).
## License
These weights are governed by the **Stability AI Community License** (see
[`LICENSE.md`](LICENSE.md)). In summary:
- **Free** for research, non-commercial, and **commercial use by organizations
and individuals with annual revenue of US $1,000,000 or less.**
- Organizations exceeding that threshold must obtain an
[Enterprise License](https://stability.ai/enterprise) from Stability AI before
commercial use of SF3D or any derivative work.
*This Stability AI Model is licensed under the Stability AI Community License,
Copyright © Stability AI Ltd. All Rights Reserved.*