Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# ONNX Image Model Collection
|
| 2 |
+
|
| 3 |
+
This collection contains ONNX versions of image upscaling and restoration models
|
| 4 |
+
intended for browser inference. Most of the models here were converted with
|
| 5 |
+
[pth2onnx-converter](https://huggingface.co/spaces/notaneimu/pth2onnx-converter).
|
| 6 |
+
|
| 7 |
+
## Built for WebGPU in the Browser
|
| 8 |
+
|
| 9 |
+
These models are primarily prepared for
|
| 10 |
+
[onnx-web-upscale](https://huggingface.co/spaces/notaneimu/onnx-web-upscale),
|
| 11 |
+
which uses ONNX Runtime Web to run image models directly in the browser with
|
| 12 |
+
WebGPU.
|
| 13 |
+
|
| 14 |
+
## What You Will Find
|
| 15 |
+
|
| 16 |
+
- Image upscalers for higher-resolution output
|
| 17 |
+
- Restoration models for JPEG cleanup, denoise, and detail recovery
|
| 18 |
+
- ONNX exports that are easier to load in browser-based pipelines
|
| 19 |
+
- Variants optimized for practical web inference where possible
|
| 20 |
+
|
| 21 |
+
## Compatibility Notes
|
| 22 |
+
|
| 23 |
+
- Most models are exported for ONNX Runtime Web usage
|
| 24 |
+
- Performance depends on the browser, GPU, driver, and model size
|
| 25 |
+
- Some larger models may require more VRAM or hit browser memory limits
|
| 26 |
+
- A few files may include manual fixes or extra optimization beyond the base
|
| 27 |
+
conversion flow
|
| 28 |
+
|
| 29 |
+
## Intended Use
|
| 30 |
+
|
| 31 |
+
Use this collection if you want to run image enhancement models locally in a
|
| 32 |
+
web app without a server-side inference stack. The main target is fast,
|
| 33 |
+
portable image processing with ONNX Runtime Web on WebGPU.
|
| 34 |
+
|
| 35 |
+
## Credits
|
| 36 |
+
|
| 37 |
+
Original model weights, architectures, and training work belong to their
|
| 38 |
+
respective authors. This collection focuses on packaging those models into a
|
| 39 |
+
browser-friendly ONNX format for easier testing and deployment.
|