YAAU is a list of model to upscale images, trained using perceptual loss.
Following fastai course on super resolution with perceptual loss (https://github.com/fastai/course-v3/blob/master/nbs/dl1/lesson7-superres.ipynb), I wondered if using custom classifier with a VGG architecture could improve the result.
This repository contains several perceptual loss model, trained on different corpus, and with different classifiers.
It also provide a generic python script to use such upscalers in a generic way, with images of any size.
|- nb: notebooks that contains the training code
| |- painting: notebooks for traning classifier and super res model for paintings
| |- danbooru: notebooks for traning classifier and super res model for danbooru
|- yauu: the python code that will upscale wholes images
|- super_res.py: the main script for inference
|- training: the folder containing the script to train your own model
two pretrained network are provided:
- super_res_painting.pth trained on paintings
- super_res_danbooru.pth trained
you can find these models here: https://mega.nz/folder/GQNDmYjQ#D6wMag7n_7LczJJtP-KQdQ
for the time being, you will need to edit super_res.py to use the right model.
python super_res.py <image_path
The result will then be in result.png
A perceptual loss model is trained in 2 stages
you need to train a classifier on the data. You will need to find a meaningful classification. For Danbooru, i trained the classifier on the rating (Q/E/S). For the painting, it was trained to classify by artist. The classifier needs to use a vgg architecture.
The upscaler is a U network with a resnet34 encoder. It is trained in a progressive way from very small image to bigger ones. You can use the scripts in yauu/training to train your own
Original picture:
upscaled with perceptual loss model:
Original:
Upscaled
you will need poetry. ./init_env.sh will install dependencies.
if you want to train the model yourself, you will need danbooru, or at least a part of it.
If you want to use a pretrained model, you can download it in the releases.
## Usage
`python ./super_res.py <image>`