search

lltcggie / waifu2x-caffe

waifu2xのCaffe版
MIT License
8.02k stars 841 forks source link

Model differences? #144

Closed FlamingMangos closed 5 years ago

FlamingMangos commented 5 years ago

I'm really confused with the large numbers of models available to you. There's

2-D illust (UpRGB Model) Photo (UpPhotoModel) 2-D illust (RGB Model) Photography, Anime 2-D illust (Y model) 2-D illust (UpResNet10) 2-D illust (CUnet Model)

I want to upscale pixiv/twitter anime art but I'm not sure which one would be best for that.

f11894 commented 5 years ago

I use Google translate. The performance of CUnet model is the highest in the data that I personally examined.

Magnify only https://drive.google.com/file/d/1EU-vo5Y01ygFL0HzMuKM47fH2_OTIWW7/view https://drive.google.com/file/d/1mmQUZjl9sNfB2uidzqMyZ5v3iBxknlna/view

Denoise & Magnify https://plot.ly/~fg1189467/55/ https://plot.ly/~fg1189467/57/

Denoise only https://plot.ly/~fg1189467/59/ https://plot.ly/~fg1189467/61/

Since the image of Twitter is YUV 420 of JPG quality 85, the best result can be obtained by using denoise level 0 or 1.

lltcggie commented 5 years ago

f11894 さんに書いていただいた通り、2-D illust (CUnet Model) のDenoise & Magnifyのレベル0か1辺りがきれいに拡大できると思います。