model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
base_model: Ruicheng/moge-2-vitl-normal
|
| 4 |
+
tags:
|
| 5 |
+
- webgpu
|
| 6 |
+
- depth-estimation
|
| 7 |
+
- surface-normals
|
| 8 |
+
- monocular-depth
|
| 9 |
+
pipeline_tag: depth-estimation
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# MoGe-2 WebGPU weights
|
| 13 |
+
|
| 14 |
+
Pre-converted fp16 weights for [moge-webgpu](https://github.com/lyonsno/moge-webgpu) — a complete port of [MoGe-2](https://github.com/microsoft/MoGe) (DINOv2 ViT-Large + ConvStack decoder, `Ruicheng/moge-2-vitl-normal`) from PyTorch to pure WebGPU compute shaders. No server, no WASM, no ONNX runtime.
|
| 15 |
+
|
| 16 |
+
## Files
|
| 17 |
+
|
| 18 |
+
- `weights.bin` (~660MB) — flat fp16 binary, all model tensors concatenated for direct WebGPU buffer upload
|
| 19 |
+
- `weights.json` — tensor manifest (names, shapes, offsets)
|
| 20 |
+
|
| 21 |
+
## Usage
|
| 22 |
+
|
| 23 |
+
The [moge-webgpu](https://github.com/lyonsno/moge-webgpu) app streams these weights automatically on first load. To reproduce the conversion from the original checkpoint:
|
| 24 |
+
|
| 25 |
+
```bash
|
| 26 |
+
python tools/convert_weights.py \
|
| 27 |
+
--model Ruicheng/moge-2-vitl-normal \
|
| 28 |
+
--output public/weights.bin \
|
| 29 |
+
--dtype fp16
|
| 30 |
+
```
|
| 31 |
+
|
| 32 |
+
## License
|
| 33 |
+
|
| 34 |
+
MIT, matching upstream MoGe-2. Original model by Microsoft Research ([MoGe-2 paper](https://arxiv.org/abs/2507.02546)).
|