search

nagadomi / waifu2x

Image Super-Resolution for Anime-Style Art
http://waifu2x.udp.jp/
MIT License
27.49k stars 2.71k forks source link

Noise reduction after zooming in can not eliminate some noise #390

Open CarbonPool opened 3 years ago

CarbonPool commented 3 years ago

Purpose: The image is enlarged to 2k, and the details are preserved as much as possible Experiment: first zoom in the picture to a higher resolution using low-level noise reduction, and then zoom in and out to 2k with 3-level noise reduction

I tried some ways to keep the details, this is my solution: v2 1280x720 -> CUNet_denoise0_scale_2x(resize to 1600x900) -> CUNet_denoise3_scale_2x -> resize_to_2k

v1 1280x720 -> CUNet_denoise3_scale_2x(2k)

The image I got has some noise that cannot be eliminated. Is there a way to solve them? The reason for using denoise0 is that sometimes denoise2 has more noise than denoise0 (artifacts)

nagadomi commented 3 years ago

The best way is to train a 4x CUNet (noise_scale) directly, but currently waifu2x doesn't support 4x.

CarbonPool commented 3 years ago

The best way is to train a 4x CUNet (noise_scale) directly, but currently waifu2x doesn't support 4x.

I may not need a 4x model. A larger pixel image noise reduction zoom can preserve the details. I am puzzled but retain more noise

CarbonPool commented 3 years ago

I added detailed instructions.

nagadomi commented 3 years ago

Noise and details are essentially the same thing. The training data generated by waifu2x restricts the input image to a raw JPEG image. if the JPEG image is processed in any way (including by waifu2x), the JPEG noise will no longer be JPEG noise, but details. So the conversion process needs to be done in one shot.