Waifu2x UpCunet on Pytorch

The waifu2x website uses (as far as I can tell) the "cunet/art" class of models which can be found in Nagadomi's repo.

Those official models are written in the old lua torch, which inconvenient as support for running that kind of model is going away.

After searching (admittedly only a little), I wasn't able to find a clean pytorch converter or reimplementation. This repo is my attempt to remedy this situation.

• waifu2x_upcunet.py

is a pytorch clone of the original model architecture described here (line 575).

• convert_waifu2x_json.py

is a utility script that converts one of the weight files here into a pytorch saved model, and an exported onnx model for good measure.

Use it like:

python convert_waifu2x_json.py noise1_scale2.0x_model.json

This will generate noise1_scale2.0x_model.pt and noise1_scale2.0x_model.onnx alongside the json model.

Requirements: torch, Pillow, numpy

Exported Objects

Here are some exported versions of the 2xScale 1xNoiseReduction model:

onnx

pytorch