Batch Image Upscale

CLI: upscale.py

Just something I made for myself for fun and upscale some old images.

Uses Real-ESRGAN w/ x4plus and x4plus_anime_6B models to AI upscale images from ./import_photo and ./import_anime directories using their respective model. The output files go to ./output_photo and ./output_anime

-o, --output-path option for a custom output location.

-u, --upscale option to resample at the desired multiplier. i.e. -u 2.0 or -u 2 for 2x size. This uses Pillow's Lanczos resampling as RealESRGAN models do 4x upscaling natively and its built-in resampling uses bicubic.

Tkinter interface: gui.py

• Wrapper for upscale.py;
• Drag and drop images to the input panel;
• Tabs for Photos and Illustrations;
• Defaults to our preset directories/folders set in upscale.py;
• Right-click context menus;
• Double click to open image in your OS user default image viewer;

Simple implementation for now until I add more features to upscale.py (as mentioned in the 'To Do')

Installation

git clone https://github.com/talzahr/batch_image_upscale.git to clone this repo.

cd batch_image_upscale

pip install -r requirements-cuda.txt for CUDA supported Torch, Torchvision and Torchaudio. Uses CUDA 12.1 by default, but can modify it for a later or earlier version such as cu118

OR

pip install -r requirements-cpu.txt for the typical Torch CPU builds.

NCNN Vulkan not yet supported

At the moment this does not utilize the NCNN-Vulkan implementation of Real-ESRGAN but should be fairly easy to implement.

To Do

1. Blending feature: take original image and upsample it with bicubic/Lanczos and blend it into upscaled image for a more realistic look.
2. Bring in realesr-general-x4v3 model: May work better with photos.
3. GUI?: Why not. (COMPLETED REV. 1)