Is it possible to train a model for bigger jpeg artifacts factors?

nagadomi / waifu2x

Image Super-Resolution for Anime-Style Art

http://waifu2x.udp.jp/

MIT License

27.49k stars 2.71k forks source link

Is it possible to train a model for bigger jpeg artifacts factors? #399

Open devalexqt opened 3 years ago

devalexqt commented 3 years ago

Ho to train model with bigger jpeg compression level?

nagadomi commented 3 years ago

noise3 model is training with very low quality JPEG (jpeg quality 27 ~ 70, jpeg encode twice). https://github.com/nagadomi/waifu2x/blob/bfd7fd9060af5ce9545a0a07d4e70daeee8206a7/lib/pairwise_transform_jpeg.lua#L69-L89

devalexqt commented 3 years ago

But script not clear all pixelization from input image:

nagadomi commented 3 years ago

JPEG artifact magnified by some image processing is not a JPEG artifact. It also looks like some kind of video encoder artifact. waifu2x only supports raw JPEG artifact denoising.