Kim2091/UltraSharp
By Kim2091 · ↓ 0 · ♥ 56 · Updated 2024-08-08
What's in this repo
This page lists the files in Kim2091/UltraSharp, including 4x-UltraSharp-fp16-opset14.onnx, 4x-UltraSharp-fp16-opset15.onnx, 4x-UltraSharp-fp16-opset16.onnx and more.
- 8 ONNX models (.onnx files)
- 19 files, including 8 .onnx, 4 .pth, 2 .bin, 2 .param, 1 .safetensors, 1 .md
- Last updated 2024-08-08
ONNX files
| File | Size | Download |
|---|---|---|
| ONNX/4x-UltraSharp-fp16-opset14.onnx | — | Download |
| ONNX/4x-UltraSharp-fp16-opset15.onnx | — | Download |
| ONNX/4x-UltraSharp-fp16-opset16.onnx | — | Download |
| ONNX/4x-UltraSharp-fp16-opset17.onnx | — | Download |
| ONNX/4x-UltraSharp-fp32-opset14.onnx | — | Download |
| ONNX/4x-UltraSharp-fp32-opset15.onnx | — | Download |
| ONNX/4x-UltraSharp-fp32-opset16.onnx | — | Download |
| ONNX/4x-UltraSharp-fp32-opset17.onnx | — | Download |
How to use
These ONNX files can be loaded with ONNX Runtime. To inspect a model graph, upload the file to the Netron viewer. Download links point to the original files on Hugging Face.
README
Name: 4x-UltraSharp (config and presets included!)
License: CC BY-NC-SA 4.0
Link:
Model Architecture: ESRGAN
Scale: 4
Purpose: Any, it's universal. This model performs best on JPEG compressed images.
Iterations: 150k
batch_size: 4-8
HR_size: 128
Epoch: ~480
Dataset: So many. I used: RAW images shot by myself, SignatureEdits, AdobeMIT-5K, DIV2K, TLOK from brucethemoose, some rock/stone images from ALSA, and many images provided by @esrgan (thanks!)
Dataset_size: uh, ignore this. anywhere between 2k and 8k full size images throughout training
OTF Training Yes (custom augmentation presets)
Pretrained_Model_G: 4x-UniScale-Balanced
Description: This is my best model yet! It generates lots and lots of detail and leaves a nice texture on images. It works on most images, whether compressed or not. It does work best on JPEG compression though, as that's mostly what it was trained on. It has the ability to restore highly compressed images as well!
The model was trained with KernelGAN (thanks musl for supplying the blur kernels), noise patches, custom augmentation presets (are in with the model), and the losses: pixel, feature, cx, ssim, lpips, and fft. Mixup was used for a while, but abandedoned due to stability issues.
Gradient Clipping helped immensely with model stability throughout training.
Big thanks to musl for giving advice on how to further improve the model!
Model info
Created: 2024-01-03
Updated: 2024-08-08
View on Hugging Face