sanime2x

requires python (tested under 3.14) and:

• torch
• torchvision
• Pillow

training

• edit & run main.py for plain CNN w/ global residual link and main_resnet.py for ResNet.
• definitely download sanime2x-2k, unzip & change the path. the dataloader script expects the part_{i}/{id}.jpeg structure unchanged.
• use dataloader if you don't want to pre-load all the images into the memories (but are willing to take the speed penalty of decoding JPEG images on the fly)
• enlarge.py and enlarge_resnet.py are example code of how you can use the trained models.

trained models

you're expected to train it yourself (i don't know how good the thingy you need it to be), but trained models can be found at HuggingFase.

license

code & model is in cc0/public domain.