waifu2x
waifu2x copied to clipboard
Jpeg noise removal not working
or at least it doesn't appear to be...
It works. What's the trouble? Please hard refresh your browser (maybe CTRL+F5) and try again, as today I have updated the front-end scripts.
I have the same issue sometimes.
I like to use the site to clean up jpgs for a fan wiki I like to edit. But sometimes when I want to compare the Medium noise removal with the High, one or the other doesn't work!
Here's a image of the settings I was using:
I tried both the Convert button and Download but it makes no difference. Adding upscaling does fix it however, I would prefer to use the None option for wiki images.
How does it not work? Server error? or it is not the expected conversion result?
When I save the image after clicking either Convert or Download it is no different from the original file.
There is no Server error or anything.
Here's a comparison between the original artwork and the Waifu export with the High setting:
Notice how there's no difference.
Compare that to the Medium setting where there is a visible difference.
I think it is the same problem as below. It is not a bug in the program, but a performance issue with the machine learning model.
It is a known issue of CUNet model(defualt model). level0~2 overfits the JPEG quality(noise level) of the input image.
noise_level=0: Expected input image compressed with JPEG quality 85-95 noise_level=1: Expected input image compressed with JPEG quality 65-85 noise_level=2: Expected input image compressed twice with JPEG quality 27-70
If the input image is compressed once with JPEG quality 80, denoising result is better with level1 than with level2. level2 may not remove weaker noise than expected... level3 simply denoise strongly.
(level1=Medium, level2=High) related: https://github.com/nagadomi/waifu2x/issues/320#issuecomment-570095404
I am currently reworking the training code and will probably be able to fix this issue in the next few months. The level2 should have included all noise patterns lower than level2.
Sounds good! 👍Thank you!