Posted under General
Naretla said:
I've made some tests here: https://imgur.com/a/jWZt7
I like "no noise reduction" best, for most of the images. Sometimes "low noise reduction" helps smooth out curves and make lines pop out more, at the cost of some detail. I guess higher noise reduction may be more useful if the source is not so clean or crisp.
Noise reduction is most useful on images with noticeable jpeg artifacts or scan dust or similar.
I'm glad people are talking about this because I am almost the sole uploader (aside for some others) for the recent of many upscales. One contributor (I'll keep anonymous) reached out to me about it and asked me to use my upscale as a parent because they believed it was higher quality and resolution, but I tried to explain to them why it shouldn't be, I think they missed that it was really a upscale. Here is what I told them:
...The recent series of child posts I've uploaded are upscaled versions made from the original images I've cited as parents. I use the source artwork with a rather small resolution and enlarge them with noise reduction. Unfortunately, the technique and software I use is not flawless, it's algorithms try to make the image larger while keeping it clear (not just blurring it), but unavoidably the upscaled versions receive quality imperfections that do not make them perfect. Thus, while the image I posted is higher in resolution, the more important criteria is image quality, which is as high as it gets from the original work. I tag all upscaled images with the 'upscaled' tag and edit them as child's of the parent post I originally used to create them. My versions are strictly inferior in quality. ...
One tricky part is by what do we judge quality by? The lack of artifacts? Authentic to the artist's creation? How can we know that between image X and Y of the same artwork which one is more truthful to the artist's intention considering the difference extremely minute? I did not tell them that I indeed was mostly using Waifu2x's engine to create these because I did not want to get word out and cause a flood of upscales. I upscaled many images and spent a lot of time choosing which images and what parameters would produce the best results for danbooru. Alas, I regret some of my uploads because I agree that danbooru does not need to be inflated with different versions of posts that people can create themselves with the third-party tool. However, I also agree with what CodeKyuubi mentioned earlier:
CodeKyuubi said:
I think there is meaning in uploading good upscales of smaller images.
Please look at post #2044134. I picked what I considered to be a very good original picture of low resolution that I could not find any better quality versions of. I spent some time to upscale it because I believed it would be very popular. It's a fantastic picture that people could use for wallpapers or whatever if only it was in a decent resolution. I uploaded this upscale as a child. It turns out it was indeed popular, and someone else had changed it to a parent. It is my only upscaled post I've posted that is a parent. As I said before, the upscales should be strictly inferior to the original, but the problem is the original in this case has numerous problem such as artifacts or jagged lines. I spent a lot of time on this post particularly comparing the two and I think people can agree that the smooth, solid lines for the character's body (hair outline is particularly awful in the original) is a major improvement. I believe this so strongly that I believe that the few granular details lost in the upscale will not only go unnoticed, but that in general the upscale is more valuable than the original! [Homage to Kaiki].
As for the policy for the future. If I had to imagine danbooru in three years without any special regulation for upscales, I dislike it. Too much redundancy and I know I would disagree with a lot of what might get approved. I would think bloat, bloat, bloat. I love the option to integrate an option to upscale like S1eth suggested, but it may not be a feature with enough demand to warrant adding it. Looking back at it, I could've uploaded much less upscales than I did--being really, really judicious on when to upload an upscale, instead of also uploading images I simply personally liked such as post #2035308. I haven't uploaded a new picture for a while because my current standards for it are high enough that it takes some time commitment to look through thousands of images and compare tens of versions of an image before I decide "yep, let's upload this upscale." Again, there are some upscales I uploaded I think are higher quality, but not necessary for me to upload.
I do not think Waifu2x should be tagged as there is no way to identify the upscaling algorithm used by looking at the finished product. For example, can anyone tell me which upscaled posts used waifu2x or did not? Only the uploader. Instead, it may be better to continue using the 'upscaled' tag and in its wiki page list Waifu2x as an effective upscaling website anyone can use.
Finally, if upscaling is going to remain acceptable, the criteria for accepting upscales should be higher than for a new post. They should not only be well done (not removing important lines, etc.), but they have to show significant improvements over the original, which also should've been a good image to begin by the standards of the approver. For example, I think my earlier Hibike! Euphonium upload is one of the best upscales here, while due to the nature of the original, this post# 2047422 is not that great. It will take people some time to learn what posts upscale well and which do not. The conversation should recognize both the need for regulation/awareness of the topic as well as really how much of an improvement in quality some upscales can bring.
EDIT: Added a bit, removed a bit.
Updated
Kikimaru said:
-1 to approving upscaled images.
They are not the original artist's vision/version and as such should not have a place here.
By all means, use waifu2x for personal collections, but images should be sourced & sized correctly.
They are not the artist's version, but how can you claim that they are not the artist's vision when, for all intents and purposes, it is the same image?
We might as well disallow recolors, colorizes, and scan-only images then, because the artist never intended for users to see the images in color/digitally.
CodeKyuubi said:
They are not the artist's version, but how can you claim that they are not the artist's vision when, for all intents and purposes, it is the same image?
We might as well disallow recolors, colorizes, and scan-only images then, because the artist never intended for users to see the images in color/digitally.
Scans only exist because the digital versions are not available.
Recolors are art by the person who colored it manually, with effort. (rather than having software do it for you. Not everyone can just recreate a recolor in a few seconds)
Consider this post #792796 and its child post #1500510. (If I'm not mistaken) the child is a beautifully well rendered vector version that stays true to the original scan. Both the parent and child are fantastic, but there are dramatic enough differences that I believe that without a doubt the child deserves a place. Creating a vector takes a lot of courage to decide how to portray it. For example in the vector version of Kuroneko's lower left chin uses a soft edge while in the original it looks like a hard edge. There are other areas which are shaded differently (e.g. right wrist), but the total result is still a very high quality image that has features the original lacks (background less, clean shades, clean lines, white background, resolution, etc.).
My argument is that not all upscales are just herp-derp big resolution versions of the original. Many of them clean artifacts, smooth jagged lines, make the lines cleaner, etc. The problem is certain images upscale poorly and look a tad blurry, sacrificing too much for just higher resolution. Now where do we draw the line between good upscales and poor ones? That's tricky, and when there are who-knows how many uploads a day coming from all sorts of sources, how does an approver know at first glance if this image is the original "artist's vision," or if this child deserves a spot and that one doesn't? Well, you can make a judgement call, and it's subjective, and that's how a lot of things on danbooru work.
The difference, to me, is the effort put in. Feeding every image ever into an algorithm and vomiting out upscales that look pretty much like the originals with some filters applied is cheap and doesn't contribute much other than them just being bigger. If the artist wanted the picture to be bigger, they'd probably have made it that way to begin with and upscaling them via "machine" at best adds nothing but artificial size and at worst damages the image.
In that example the upscale is done at least partially by hand. It's basically an art of its own (though I'm sure some people consider it "tracing" and thus "cheating," though less so due to it not trying to hide it by changing the character/mirroring/etc.). Effort was put in and a not-entirely-exact-copy was put out. It had a purpose (namely removing the background) and was checked, edited, and corrected by the person doing it, whereas the waifu2x images we're talking about here seem to be "paste into thing, push butan, receive big one, upload."
As to detecting and dealing with these machine upscales, I'm not sure there's a reliable way to do it. However they should be child posts, and I personally don't think they should be approved. This ties in a bit with the other recent discussion about using thumbnails in the mod queue; one certainly can't tell an upscale from there.
snip
Kakisho said:
For example, can anyone tell me which upscaled posts used waifu2x or did not? Only the uploader.
Test it manually. I noticed that the resolution for post #2055902 (upscale) is exactly double the resolution of post #1406374 and remembered that waifu2x was around. I ran the original with no noise reduction and 2x and got the same image.
S1eth said:
Scans only exist because the digital versions are not available.
Yes, because the artist intentionally withheld the image, because he/she wanted the viewer to buy a book to see the art, or pay some sum.
Regardless of how the discussion goes, I'll go along with what the majority decides. I personally wouldn't upload an upscale like this, though I sometimes do similar things with scans (Stitching together pieces of different scans to create a higher resolution, as clean as I can make it image (like post #1742899)). I'm just here to offer my opinions on the matter.
I just feel like, if someone puts in the effort to blow up an image tagged with lowres and clean it and sharpen the lines while retaining as much of the original's quality as possible, there is meaning to having upscales, because I'd say many users are wont to have them at a larger resolution to better appreciate the image and the details within.
While I do believe we should discourage the upload of upscaled images, but people are becoming unreasonable in what they're asking for. It is unreasonable to say not approve these images when approvers may not be able to tell if an image is or isn't an upscale without conducting timely tests. Those tests may not even necessarily give the approver confirmation that it is an upscale and so all it results in guessing. Furthermore tests can only even be performed if we have another version to go by, which isn't guaranteed.
We should incorporate this kind of upscaling into the site in some form though.
CodeKyuubi said:
I just feel like, if someone puts in the effort to blow up an image tagged with lowres and clean it and sharpen the lines while retaining as much of the original's quality as possible, there is meaning to having upscales, because I'd say many users are wont to have them at a larger resolution to better appreciate the image and the details within.
I feel that if an image is of good quality and it took a notable effort to produce, even if it's 'only' an upscale, then making it available would offer a significant benefit. But even though waifu2x can produce good images, the effort required is still too low: it is trivial to do it yourself, as long as the public demo remains up. (It might be neat to integrate waifu2x into the site somehow, but that's another matter.)
(Hypothetically, if the author of waifu2x had kept it private and only used it to enhance select images, then the individual products would be more valuable to us.)
Of course, that doesn't necessarily mean an image is disqualified if waifu2x was involved in its creation, just that the process can't be as simple as ticking a couple options: if other tools were also used, or a private instance of waifu2x running on more advanced settings, there might be an argument for approving it.
EDIT:
NWF_Renim said:
We should incorporate this kind of upscaling into the site in some form though.
That would be pretty cool.
Updated
CodeKyuubi said:
I just feel like, if someone puts in the effort to blow up an image tagged with lowres and clean it and sharpen the lines while retaining as much of the original's quality as possible, there is meaning to having upscales
Yes, but that is not what this topic is about. This is about no-effort-one-button-press upscales that anyone can do at any time on their own without the help of an artist.
A 2-page image scan is an altered version from the original. When you're stitching them together and cleaning them up, you're bringing the altered version closer to the original, which is good. Because we want the original or at least a version with as little differences as possible.
psich said:
Test it manually. I noticed that the resolution for post #2055902 (upscale) is exactly double the resolution of post #1406374 and remembered that waifu2x was around. I ran the original with no noise reduction and 2x and got the same image.
The resulting image depends on the images used to teach the algorithm, doesn't it? That would mean that the same MD5 hash is not guaranteed for 2 upscales of the same image.
Testing definitely doesn't seem like an option there, asking approvers to add another step and running images through the site seems like a colossal waste of time. Personally I'd just be wary of unsourced upscales, especially where the original has a source. The example above (post #2055902), for instance, has no source, while it's parent is on the artist's pixiv. That would throw up a flag telling me it's user made as is, even if I hadn't been told.
NWF_Renim said:
We should incorporate this kind of upscaling into the site in some form though.
I agree. That would add a very useful feature to the site, while also making users a lot less likely to upload upscales if the originals have already been uploaded: why upload the upscale when the site can just make it? Two birds with one stone.
Hell, I bet the algorithm could get even better if it was trained here. There aren't many places around that have that many well-organized high-quality anime-style images to choose from. Not sure how feasible that would be technically though.
What about using waifu2x for noise reduction on images where the best available version has jpeg_artifacts? Twitter images for example get heavily recompressed. It feels like photoshopping to me, but it also makes some images look much better.
lkjh098 said:
What about using waifu2x for noise reduction on images where the best available version has jpeg_artifacts? Twitter images for example get heavily recompressed. It feels like photoshopping to me, but it also makes some images look much better.
You would likely have to go into the image with photoshop anyway to add masks, so that you don't lose too much detail.
CodeKyuubi said:
You would likely have to go into the image with photoshop anyway to add masks, so that you don't lose too much detail.
I've had some practice at fixing artifacts with photoshop, and waifu2x is much better at preserving detail than I am. The main problem seems to be it misses some low-frequency artifacts.
I've found it good for flat images with clean lineart, but with any scans it was really unhappy doing anything with them, especially anything with a texture of paper. That said, it definitely works better than nearly all the upscalers I've seen.
Maybe we can discuss a minimum resolution for this, because I think there's a certain natural cut-off point when the image just doesn't look right, and gets too many artifacts.
