Close
This post was deleted. Reason: Upscaled. MD5: 6a7a59cc3a860fa17101e992b5da328a
This post belongs to a parent post.


Edit | Respond

It appears to be upscale.
Can you provide a source for this picture?
It was still upscaled.
BTW
I tried a new upscaling algorithm using picture post #1005817.
https://imgur.com/8v1i0mM
If the quality improves so much, it may be harder to discriminate upscaling in the future.
hiroimo2 said:
It was still upscaled.
BTW
I tried a new upscaling algorithm using picture post #1005817.
https://imgur.com/8v1i0mM
If the quality improves so much, it may be harder to discriminate upscaling in the future.
It seems to work well, what kind of algorithm is it?
whitespace1 said:
It seems to work well, what kind of algorithm is it?
It's Real-ESRGAN.
The differences between this and waifu2x can be found in the URL below.
https://github.com/xinntao/Real-ESRGAN/blob/master/docs/anime_model.md
hiroimo2 said:
It's Real-ESRGAN.
The differences between this and waifu2x can be found in the URL below.
https://github.com/xinntao/Real-ESRGAN/blob/master/docs/anime_model.md
Thanks hiroimo ❤
Uh, not sure using the worst compressed images as inputs is the best way to compare results.

Don't get me wrong, artifact (compression or else) cleaning is a very useful aspect of these AI algorithms, but most of people start with an small but clean image.

They should at least show the result from these use cases as well.

And it looks me that if you just apply a boring sharpen filter after waifu2x (which produces blurry result because.. the input *is* blurry af), the result would be even closer.

Anyway, I think this algorithm works better (probably the intended use anyway) for upscaling videos like DVD etc. to HD.
I also think this is an engine for anime DVDs.
However, I noted the second mp4 in samples.
This video shows a before and after comparison for an illustration. The blurring is surprisingly improved in full screen. This effect will greatly improve the quality when the picture is enlarged.