Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

anime4K paper #105

Open
yushengnan1 opened this issue Oct 9, 2020 · 3 comments
Open

anime4K paper #105

yushengnan1 opened this issue Oct 9, 2020 · 3 comments

Comments

@yushengnan1
Copy link

Can you share a link to anime4k's paper? Where can I find it? Can you share the link with me?

@topin89
Copy link

topin89 commented Oct 29, 2021

If there is no paper, can you write a short description how it's that much faster than ESRGAN with without visible quality loss.
Really, how? Is it because of some layer size reduction? Is it because ESRGAN is unoptimized pytorch script and Anime4K is by-hand optimized shader? All of the above, something I can't think of? Is it a himitsu? (himitsu means secret)

@bloc97
Copy link
Owner

bloc97 commented Oct 29, 2021

The Anime4K shader is not a GAN, it was trained using pixel-wise loss. If I remember correctly recent SRGANs have 10M+ parameters while the optimized smaller Anime4K shaders have 1.6k parameters. GANs can implicitly learn to restore some unseen degradations by learning from a discriminator (the optimum in a zero-sum game optimization will be much more robust to degradation compared to a standard MSE optimization). Instead of using a discriminator, Anime4K was trained on a special dataset that mimics the degradation of 1080p anime.

@bloc97
Copy link
Owner

bloc97 commented Oct 29, 2021

This is why if you try Anime4K on very old anime it does not look as good compared to 1080p anime. It is not a replacement for SRGANs, but simply an alternative to watch 1080p anime on 4K screen in real time.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants