SuperResolutionWaifuNN

https://user-images.githubusercontent.com/47704357/153986586-6b53fed9-4b5d-4a3e-9c07-202be5b1d1c0.mp4

I wanted to build my own implementation of waifu2x using Keras and TensorFlow, but I ended up using a slightly different architecture.

Image dataset colected using PixivUtil2.

Flutter implemetation using tflite srwnn_mobile.

Training

The model was trained on about a thousand images. Better quality of the gif: https://youtu.be/GCOe5kMTM5E

Test (using 8k+ images for training)

Input/Output

Tests 1

The first image is the low ressolution input, the second is the generated image and the third is the original high resolution image.

Tests 2

Denoise test

Deblur test

Test (using 1.5k images for training)

Tests

The first image is the low ressolution input, the second is the generated image and the third is the original high resolution image.

Test 1

Test 2

Test 3

Test 4

Test 5

Test 6

Test 7

Test using non anime picture