[X] My Assistant

Then either you can re-process the same original files, upload them as a new gallery and call it superior version, or someone else can do the same with your original files.

This is already not permitted.

We should probably come up with a tag for these kinds of upscales. Yes, it will be on the honor system as to whether it was upscaled with a nn or not.

We cannot tag them as there is no objective means to detect them.

This is already not permitted.

Not sure if I got it right. Referring to this:

Not really, we would allow only a single "upscale" gallery (the best one).

To me this means a case where there are more than one enhanced version of some gallery. Which means it's ok to upload another modified version of it as long as it's better than the last best one.

We cannot tag them as there is no objective means to detect them.

There are certain defects that help detect processed images in many cases, just need a good eye. Also, with digital galleries you can look up information about the work on usual portals and find the huge difference in file size. Sometimes they put actual image resolution in description.

Also, I think those two galleries which were mentioned are not even the output of waifu2x. I got better result by simply resizing one of original images.

There are certain defects that help detect processed images in many cases, just need a good eye. Also, with digital galleries you can look up information about the work on usual portals and find the huge difference in file size. Sometimes they put actual image resolution in description.

Also, I think those two galleries which were mentioned are not even the output of waifu2x. I got better result by simply resizing one of original images.

To be fair I'm kind of tired of repeating this same quote:

Took me too long to finish my beer, I literally dozed off about this thread (that's an internal administration joke )

To extend on Joe's post above we are well aware of waifu2x and well aware on how it performs its inner workings. And we can certainly say that waifu2x is simply awesome. As for the relevant rules have a look at the definition of superior versions on the wiki:
https://ehwiki.org/wiki/Expunging#Superior_Versions

And the stance of the admin/mod team is that waifu2x is a valid form of retouching a scan. All other relevant rules for superior versions (on that same page) also remain intact.

In summary:

1. A gallery retouched by waifu2x is a valid gallery
2. A gallery retouched by waifu2x does NOT constitute a superior version to an existing non-retouched gallery
3. A non-retouched gallery does NOT constitute a superior version to an existing gallery retouched by waifu2x, unless it is a digital release in which case any form of retouching would not be consistent with "as released by the source".

waifu2x is a deep convolutional neural network. It does not perform a geometric (binomial or cubic) upscale of an image. Instead it bends the information contained in the image towards a more anime-like appearance. The fact that it is capable of producing bigger (in terms of resolution) is not relevant. The reason that a geometric upscale is bad is because it reduces the amount of information per square inch (or any other measure you may use) of an image. waifu2x does not reduce the amount of information per square inch, it fill in information (guesses) based on the networks knowledge about anime. This produces a slight loss of information of the original image whilst improving the overall consistency of the image by filling in information that seems more relevant.

By that evaluation waifu2x performs the same work as (good) grayscale levelling, loss of information which is weighted by image consistency.

This thread is now sticky, to ensure that everyone relevant sees it. (And to work as a precedence to similar software)

Back to your own quote (post, whatever):

So no. (1) is just a quote out of context.
And for (2) I have an MSc in neural nets, trust me on that one. (And the thread already has some 6-7 posts on the subject so am not really willing to quote myself again )

This post has been edited by blue penguin: Nov 21 2017, 07:50

To be fair I'm kind of tired of repeating this same quote:

Sorry but I don't see any conflicts with what I've said. It's possible to distinguish waifu2x-based upscales in many cases, and some conditions increase the chances in my experience:
- the original image is not very tiny (bigger than 800x600 or so).
- the original image has some JPEG artifacts.
- the scaling factor is 2.0 or higher.

One thing that is going to be really difficult to detect in most cases is when you apply denoise without upscaling.
I can try demonstrating some defects that waifu2x produces in magnified images if someone has worthy samples they are interested in.

So no. (1) is just a quote out of context.

Do you mean that after one reconstructed gallery is posted, no more reconstructed galleries are allowed even if they look better? I mean not a case where x3 upscale may look better than x2 one (it will most likely not), but a case where an image will be produced with different or improved method (not waifu2x) at similar scale factor.

This post has been edited by genl: Nov 21 2017, 17:12

We cannot tag them as there is no objective means to detect them.

There are certain defects that help detect processed images in many cases, just need a good eye. Also, with digital galleries you can look up information about the work on usual portals and find the huge difference in file size. Sometimes they put actual image resolution in description.

sorry to butt in, but... even if this was possible (i don't know since it's not my cup of tea, but let's think for a moment it is, just because) and, saying, a "waifued" tag appears, wouldn't it be a bit dangerous and/or prone to abuses? at least this is what comes into my mind, judging from previous discussions about this matter...

This post has been edited by Scremaz: Nov 21 2017, 20:16

This debate about flagging such galleries, personally, it's not something that should be a problem in the first place.

The rule should've been always "only original images allowed" (it's quite easy to see when a pic is even remotely recompressed, even without making any edit or adding any filter whatsoever [unless you're using bitmap pics ]), full stop.

All of this is a consequence of making exceptions to a rule. Everything accepted as long as it's the first version ever uploaded. Now enjoy them. "Fortunately" the shit ones are quite recognizable even without a tag flagging them.

As "them" I'm talking about all these galleries that are """waifu0.5x""" (or waifu2x thumbnails) on purpose, of course.

And by the way, a reworked version that [partially or totally] fixes the horrible aspect of some grained/blurred ones, should not only be accepted as a replacement, but also praised and appreciated; if "better than nothing" is better than nothing, "a better version of better than nothing" is definitely better compared to "better than nothing".
Of course if you try to abuse this without actually improving the edited gallery, it mustn't be allowed. But expunge petitions exist for that. You expunge either the dupe or the old one. Same as before. With the only difference that right now the second gallery is expunged no matter what (unless it's not an edit but an alternative better scan, but this is not the problem we're talking about right now, and it's also a way rarer occurence, aka the majority of the bad galleries are not [and probably won't ever be] replaced with alternative scans). So right now the amount of work for the expungers is the same if someone does a replacement, but the result is worse.

This post has been edited by End Of All Hope: Nov 21 2017, 21:22

Okay, once again, here is the official stance on the matter.

waifu2x is awesome, so I'm certainly not going to ban it.

If you have suspicion of system abuse or feel like something is working backwards, you should see this directly with Tenboro and offer objective suggestions.

Sorry but I don't see any conflicts with what I've said. It's possible to distinguish waifu2x-based upscales in many cases, and some conditions increase the chances in my experience:
- the original image is not very tiny (bigger than 800x600 or so).
- the original image has some JPEG artifacts.
- the scaling factor is 2.0 or higher.

OK, ok, here we go.

Yes, that is a fair argument. The issue with it is the upfront assumption: that you know the source and the processed image. You can distinguish waifu's patterns if you know where to look for them and compare against the original. What you can't do, is to, having a group of images of the same thing from different sources, distinguish what is the original and waifu'ed and what are two separate originals.

We face the second case more often than any other case really. That is why the ruling on waifu is what it is.

Geometric upscales are trivial to detect because you can see the workings of the algorithm all over the place. NNs are harder.

That said, yes, if someone uses waifu and does an x12 it is trivial to detect. Waifu probably stops working well after x4, never had the computing power to make the actual graph for it though.

sorry to butt in, but... even if this was possible (i don't know since it's not my cup of tea, but let's think for a moment it is, just because) and, saying, a "waifued" tag appears, wouldn't it be a bit dangerous and/or prone to abuses? at least this is what comes into my mind, judging from previous discussions about this matter...

I'm not talking about whenever or not people should add a tag (for now at least). Only about how it's possible to discover the fact of waifu2x usage in many cases.

Yes, that is a fair argument. The issue with it is the upfront assumption: that you know the source and the processed image. You can distinguish waifu's patterns if you know where to look for them and compare against the original. What you can't do, is to, having a group of images of the same thing from different sources, distinguish what is the original and waifu'ed and what are two separate originals.

We face the second case more often than any other case really. That is why the ruling on waifu is what it is.

Geometric upscales are trivial to detect because you can see the workings of the algorithm all over the place. NNs are harder.

Not exactly. I really believe that it's possible in many cases even without having the source image. Well, if the source is digital and it's possible to find other original images from the same artist, it'll help of course. But it's not necessary in many cases.

In case with groups of images from different sources (I suppose different artists at least), wouldn't it be even more obvious? I expect that each new image with different art style would only increase chances of identifying the usage of w2x on a whole group if images.

Just a few weeks ago I identified w2x upscale usage on a digital manga gallery and uploader confirmed that in a pm (turns out there are certain b/w patterns that look just weird after w2x upscaling). DLsite page with file size information helped with assumption, but I didn't have access to original images.
In CG galleries mentioned in this thread I could see a simple upscaling (geometric as you call it), without any w2x/NN signs. And I don't understand why at least one of them is still not expunged for being an awful upscale ( http://ehentaihip.com/g/989711/bf2801f08a/ ). You can see certain pixelation. That's absolutely not how w2x upscale would look.

That said, if the source is digital and uploader applied w2x to upscale images for a gallery, it should be relatively easy to find out. When the source is a paper scan, however, I can expect 2 scenarios:
- Uploader found a low quality/resolution scan on the internet and decided to enhance it. I expect such cases to be only slightly more difficult to distinguish than with digital images. The result would still look relatively bad and most likely would lack the expected level of detail for its resolution.
- Uploader who created a scan decided to enhance it. At this point I don't have enough knowledge about usual processes, but I'd expect him to increase the scan resolution and/or use denoise feature of w2x, not its upscale feature.

In case with groups of images from different sources (I suppose different artists at least), wouldn't it be even more obvious? I expect that each new image with different art style would only increase chances of identifying the usage of w2x on a whole group if images.

Nope. The waifu2x that you get from the released bundles is already trained therefore you do not see image interference between the set of images you are processing. i.e. the batch being processed itself is irrelevant for waifu. (although it is very important in NN training alright)

- Uploader found a low quality/resolution scan on the internet and decided to enhance it. I expect such cases to be only slightly more difficult to distinguish than with digital images. The result would still look relatively bad and most likely would lack the expected level of detail for its resolution.

I would argue that x4 or x5 would like that alright. But x2 would be too difficult to figure out. That said I absolutely do not have the data to make an actual graph of this, so it is going only by guesstimates as of now.

- Uploader who created a scan decided to enhance it. At this point I don't have enough knowledge about usual processes, but I'd expect him to increase the scan resolution and/or use denoise feature of w2x, not its upscale feature.

And we definitely want to allow that, right? Note that there will be data loss in simple denoise.

It is also about the risk of killing valid retouches by being too stringent. What we did is simply a good enough compromise.

Nope. The waifu2x that you get from the released bundles is already trained therefore you do not see image interference between the set of images you are processing. i.e. the batch being processed itself is irrelevant for waifu. (although it is very important in NN training alright)

I mean that if there is more variety in a collection of images, there will be more chances for certain defects and clues to appear on specific images. It's not about a system being already trained, it's about a trained system being somewhat limited and unable to be trained further. At least during normal usage a user can't train it, from what I understand.

I would argue that x4 or x5 would like that alright. But x2 would be too difficult to figure out. That said I absolutely do not have the data to make an actual graph of this, so it is going only by guesstimates as of now.

I've used w2x to make x1.5 and x2.0 upscales on some images, b/w and color. In absolute most cases I could say that the result does not look original. Like I've said - if you have any processed image or collection of images to show, I'll look at them and try to demonstrate the clues.

And we definitely want to allow that, right? Note that there will be data loss in simple denoise.

Again, wasn't talking about whenever or not we need to allow anything. waifu2x can act as a denoise filter. I know that people who create paper scans also do all kinds of processing on images before uploading them. What I mean is that at this processing stage, I don't see any point in using waifu2x except only its denoise feature. There should be no point in magnifying scanned images for a person who is creating a quality scan. So I don't expect it to even be useful for such uploaders.

This post has been edited by genl: Nov 23 2017, 17:11

Example of waifu2x-enhanced gallery: http://ehentaihip.com/s/bfac6fa9e5/1149467-17
Here are clues marked on just one page:

At this resolution, you expect sharp symbols, but they are blurred and even feature round edges. You expect clean lines but you see inconsistent curves. You expect usual gradient but you see ripples.

This post has been edited by genl: Feb 21 2018, 06:14

What about posting a mega or gd link to the upscaled version in the comment section?

What about magnified sets like this without filters where one pixel takes up a 10x10 pixel space?

http://ehentaihip.com/g/1270046/2d2b62c3e2/

http://ehentaihip.com/g/1270046/2d2b62c3e2/

Oh god, that's some really ugly upscale.

If anyone can find the non-upscaled version it should be expunged on spot.

If anyone can find the non-upscaled version it should be expunged on spot.

I tried to replace the gallery but the uploader gave an error "file was created by this gallery, replacing with the original" and it replaced my 9MB gallery with the 200MB version.

Did you just attempt to replace with the versions made by E-H's downscaler...?

To extend on Joe's post above we are well aware of waifu2x and well aware of how it performs its inner workings. And we can certainly say that waifu2x is simply awesome. As for the relevant rules have a look at the definition of superior versions on the wiki:
https://ehwiki.org/wiki/Expunging#Superior_Versions

And the stance of the admin/mod team is that waifu2x is a valid form of retouching a scan. All other relevant rules for superior versions (on that same page) also remain intact.

In summary:

1. A gallery retouched by waifu2x is a valid gallery
2. A gallery retouched by waifu2x does NOT constitute a superior version to an existing non-retouched gallery
3. A non-retouched gallery does NOT represent the first version to a current gallery retouched by waifu2x unless it is a digital release in which case any form of retouching would not be consistent with "as released by the source."

Does it mean that I can publish my uploads? Since most of my uploads are enlarged using Waifu2X

This post has been edited by IndonBullshit: May 28 2019, 14:02