Posted under General
This topic has been locked.
Benit149 said:
I'm looking around for it in my mock upload, which I'm using to experiment with looking for the metadata, not intended for actual upload -> https://danbooru.donmai.us/uploads/6296101
I still don't understand where in the Upload page I'm supposed to be looking, particularly if I haven't uploaded the post.
Go to Posts>Upload>My Uploads>Unposted and click on the image you want to check after submitting the url but before actually uploading it. You'll be taken to this.
blindVigil said:
Go to Posts>Upload>My Uploads>Unposted and click on the image you want to check after submitting the url but before actually uploading it. You'll be taken to this.
That only takes you to the upload page for the post in question, only media assets from other users direct you to the media asset page.
blindVigil said:
Go to Posts>Upload>My Uploads>Unposted and click on the image you want to check after submitting the url but before actually uploading it. You'll be taken to this.
Guys, I am seriously following your steps exactly as you say, and I'm still being taken to the Upload page for my mock upload.
That only takes you to the upload page for the post in question, only media assets from other users direct you to the media asset page.
That must be the problem. I can look at other users' metadata, but I can't look at my own before completing an upload? I fail to see how that makes sense.
Updated
Benit149 said:
Guys, I am seriously following your steps exactly as you say, and I'm still being taken to the Upload page for my mock upload.
That must be the problem. I can look at other users' metadata, but I can't look at my own?
For now you can use this script I made a while ago, it adds media asset links under the image on the upload page and below your unposted uploads on the overview page.
// ==UserScript==
// @name Danbooru uploads link media assets
// @match *://*.donmai.us/uploads
// @match *://*.donmai.us/users/*/uploads
// @match *://*.donmai.us/uploads/*
// @version 0.4
// @author Nameless Contributor
// @description Add public media asset links below unposted uploads
// ==/UserScript==
'use strict';
const assetLink = (id, div) => `
${div ? '<div class="leading-none">' : ''}<a class="leading-none" href="/media_assets/${id}">media asset #${id}</a>${div ? '</div>' : ''}
`;
fetch(`${location.href}.json?only=media_asset_id,upload_media_assets[media_asset_id]`)
.then((response) => response.json())
.then((json) => {
if (location.href.match(/\/\d+$/)) {
const id = json.media_asset_id ? json.media_asset_id : json.upload_media_assets[0].media_asset_id;
document.querySelector('#upload-image p').insertAdjacentHTML('beforeend', assetLink(id));
} else {
for (const [i, preview] of document.querySelectorAll('.media-asset-preview').entries()) {
if (preview.getAttribute('data-is-posted') === 'false' && !(preview.getAttribute('data-media-asset-count') > 1)) {
const id = json[i].media_asset_id ? json[i].media_asset_id : json[i].upload_media_assets[0].media_asset_id;
preview.querySelector(':scope > .text-center').innerHTML = assetLink(id, true);
}
}
}
});
Benit149 said:
Is there a way to tell if an image is AI-generated and the artist doesn't use the appropriate tags, like what happened to me when I uploaded post #5749360? I've been doing my best to filter out any AI-related tags on Pixiv for my searches, but clearly it's not as solid a system as I would like. I know just looking at the image and spotting the telltale signs of an AI-generated illustration is somewhat helpful, along with the guidelines described in the ai-generated wiki, but it's still so troublesome when it looks so samey and could masquerade as an artist's actual style.
Just for refrence, here's a quick redline of the obvious stuff to me.
Edit: So new question. What should we do about people that have just signed up to dump AI art, and their own genrations at that. user #951290 and user #953730 for example?
Updated
AI Generated Art Should Be Allowed. It solves a certain issue. Some characters that deserved to have more can have those needs fulfilled by the AI. A Sexy girl with low amount of art is just very saddening. 2nd Reason, from what Ive seen, AI Lewds definitely are not as creative as us Real Humans. Alot of it seems like nothing too extraordinary. 3, It seems like It can really only trained and basically make stuff off of stuff thats already been done. Is there really any reason to stop something thats just redoing the same stuff we've already seen? It doesn't really seem like any sort of threat at all tbh.
iLewdWaifus said:
AI Generated Art Should Be Allowed. It solves a certain issue. Some characters that deserved to have more can have those needs fulfilled by the AI. A Sexy girl with low amount of art is just very saddening. 2nd Reason, from what Ive seen, AI Lewds definitely are not as creative as us Real Humans. Alot of it seems like nothing too extraordinary. 3, It seems like It can really only trained and basically make stuff off of stuff thats already been done. Is there really any reason to stop something thats just redoing the same stuff we've already seen? It doesn't really seem like any sort of threat at all tbh.
Danbooru is not a limitless storage space for garbage. Your first reason isn't a valid one, and the other two are better reasons not to allow it. We turn away mediocre art made by actual humans daily, why would we then turn around and say mediocre art shat out by a computer that doesn't even know what art is and can't even fathom the simple concept that humans don't have three legs is acceptable?
You can just use an online tool to check for metadata, too. I found good results with this site
Updated
iLewdWaifus said:
AI Generated Art Should Be Allowed. It solves a certain issue. Some characters that deserved to have more can have those needs fulfilled by the AI. A Sexy girl with low amount of art is just very saddening. 2nd Reason, from what Ive seen, AI Lewds definitely are not as creative as us Real Humans. Alot of it seems like nothing too extraordinary. 3, It seems like It can really only trained and basically make stuff off of stuff thats already been done. Is there really any reason to stop something thats just redoing the same stuff we've already seen? It doesn't really seem like any sort of threat at all tbh.
Don’t mean to dogpile or be an asshole, but no. We should absolutely not allow it. This site isn’t just for jacking it, it’s supposed to host art. You can get into philosophical discussions all you want, but it seems the vast majority of users and mods don’t see AI content as art. Even if fighting waves of AI “art” is impossible and will end in failure, might as well try to stop it while it’s still possible. I’d personally like to enjoy this website while I still can.
blindVigil said:
Danbooru is not a limitless storage space for garbage. Your first reason isn't a valid one, and the other two are better reasons not to allow it. We turn away mediocre art made by actual humans daily, why would we then turn around and say mediocre art shat out by a computer that doesn't even know what art is and can't even fathom the simple concept that humans don't have three legs is acceptable?
Because approvers can be as infallible as the art they claim is trash. It's not even an objective argument like you think is it. There are sometimes when submissions of well-established artists whose piece is rejected when the piece isn't really not different from all the approved ones. There's no reason for that to be a thing even but here we are. It's even worse when they don't even have the balls to give a reason. A critic with a trained eye is one thing, a gatekeeper is another entirely. The line between to those two worlds gets thinner as one reads this thread.
I don't agree with AI-art but my grievances, and the grievances others, aren't valid reasons if the art is good enough. The only reason people against AI-generated art can even make their case is because it isn't at some basic level of quality, which varies person to person. That's it. At some point, AI capabilities will surpass it's obvious imperfections. When that happens, there won't be a quality argument any more. However, since we aren't there now (and even that is debatable), we have this interim argument on the validity of AI-generated art. The validity argument is that AI-art is just theft has real merit, but that's a consensus ultimately irrelevant to Danbooru IMO. Most of us aren't artists at the end of the day.
I honest-to-god love AI art. I love it with all its current errors, I love it with what it gets right. I can't wait 'til the technology becomes advanced enough to be able to produce flawless highres illustrations consistently. I find it fascinating.
But I still believe it should not be on Danbooru. Setting quality matters aside - looking at the progress between last year's early November's sensation, NeuralBlender, and current capabilities of MidJourney or NovelAI, we can only expect the quality to rise - its biggest issue is how mass-producable it is/will become. A human artist can take hours creating something for themself, for a commission you can expect to be waiting days if not weeks. AIgen will (likely) only get better and faster. If I recall correctly, current NovelAI takes seconds to create an image. At some point, AI art would simply flood the site - why try to "hunt" for a good human artwork of something you like when you can just type in your waifu's name and/or characteristics you want to find her with?
To prevent that potential flood, I think keeping AI art from Danbooru is a good move. And after all, Danbooru isn't the only place where people can share their AI creations - Reddit communities, other boorus like AIBooru... They all will still exist even with the ban.
This argument is pointless for one very simple reason: Quality has nothing to do with why it was banned here.
It got banned because the makers of the AI decided to say they were affiliated with this site, and as a result caused a wave of artists to demand removal. The AI ban is a reactionary strike back, not some decision to side with the issue at hand in one manner or another.
Even if it were impossible to find all of the secret tells, even if it managed to create something perfect, that wouldn't change that it would be banned here. Thankfully, with the existence of the AI-booru, there really isn't a point of trying to campaign to change that either. Just add that link to your daily visits and it won't really be much different than if you searched for something generated.
That's not true. From the first post that was tagged as ai-generated, that is, post #5540274, there has been constant discussion regarding whether or not it should be allowed, just not all of it was on the forums. NovelAI just significantly hastened the decisions.
Nameless_Contributor said:
For now you can use this script I made a while ago, it adds media asset links under the image on the upload page and below your unposted uploads on the overview page.
Awesome. Thank you so much. This should make filtering AI-generated art a thousand times more easier on Danbooru, since I'm already doing what I can on Pixiv.
Just saw that AI-generated art and the backlash from traditional artists made mainstream news; of course in Western media it's focused on Western artists right now, but this is definitely something that's going to be a big issue far beyond the fallout Danbooru is suffering (like if game devs just start relying on AI art to avoid having to pay artists)
I was asked my opinion of the Danbooru AI-art temp ban, as someone who's been doing anime generation since 2015 but is not really part of the Danbooru community; I think it's fine—good, even.
I don't think it's cowardly to not define what's acceptable or when it'll be allowed, and punt to later.
This is for a variety of reasons, but mostly: the loss from a temporary ban is not large because the benefits are small and it can be made up later, while the downsides are large (even excluding any legal issues).
1. It is unimportant for Danbooru to curate AI art. There is still not much of it (albeit rapidly increasing), and most of it is still bad.
2. Images generated now are worse than they look. You will come to regret uploading many of the images which would be uploaded right now. Aside from the obvious artifacts like hands (hands are a problem I have struggled with for years—now you all know my pain), even 'good' samples are going to look worse in a year or two. There is, with every leap in technology, whether daguerrotypes or Lumiere moving-pictures of trains or ProGAN human faces, a honeymoon period where people are justly astonished at its realism; but as time passes, everyone sees enough of it to become sensitized to all its flaws. Consider how laughable CGI effects in Hollywood movies from the '80s look, or how terrible early computer graphics in manga & anime look now. (I read through the original Ghost in the Shell manga recently; they look terrible now, as much as they awed readers in the '90s.)
Further, use of generative models always improves considerably as time passes, as people learn how to prompt them, develop libraries of prompts, discover new prompts (early on, prompts are usually terribly bland, boring, and restricted—people just don't know what the model can do, and have too little imagination to discover it all immediately. We are still figuring out all the crazy things GPT-3 can do), apply new methods like 'negative prompts' (not even a year old, IIRC) or textual-inversion/DreamBooth etc. You may be impressed by current samples, but soon you will be looking at better samples and finding the old ones to be garbage. You will flinch at the small images, upscaling blurs, square aspect ratio (totally unnecessary, there's always been many ways to generate variable aspect ratio images), and other limitations now. Similarly, with video games, the final games on a platform tend to be the most visually spectacular, because that is when the game developers have learned to use the hardware to its fullest.
I would also remind everyone that deep learning only gets better: while Stable Diffusion is driving this conversation, SD is well below the publicly-known SOTA of image generation (there are surely even better models than Imagen & Parti cooking away, and who knows about GPT-4?), and the research frontier is already shifting to video. You shouldn't fixate on SD, but be thinking about images benefiting from video/3D-modeling knowledge in models trained with 100x the compute of Image/Parti and refined with years of tricks & optimizations. Because that's where things will be not terribly long from now...
3. Images added to the database will be a burden forever. It is better to have a bright line rule than either let bad images get in (as curation crumbles under the workload), sit around indefinitely, or be purged inconsistently to the understandable anger of everyone who invested any effort in finding/uploading/tagging them.
4. The proper way to handle AI images is still up in the air. Any set of tags or rules developed in haste, pressured by uploads of possibly hundreds or thousands of images a day, may be severely flawed, and overly tailored to the NovelAI SD model. The future is one of many different models and workflows for years or decades to come, and no one can foresee it. So, why not pause it a while, and see how things shake out? After all, look how much has changed in just the past 2 months! You can't be sure where things will be 2 months from now, much less 2 years; but the consequences of any mistakes in screwing with the tags or letting in the wrong images or modifying the software will still be with you. Why not let other sites like Pixiv and AIbooru figure things out first, or serve as object lessons in how not to handle AI images?
Given the rapid increase in number of generated samples, a mistake in curation could lead to a lot of images being uploaded. It's hard to get statistics, but given that Pixiv search results for 'NovelAI' are already yielding >30k hits. I don't think there is much risk of Danbooru becoming majority AI in the next year, but in a few years, it will become possible to generate hundreds of thousands of very high-quality diverse auto-tagged images automatically and a few dedicated humans screen them for upload, and simply swamp Danbooru while following the letter & spirit of the rules; before that thing happens, it would be good for the rules to be correct so that is a good thing and not a bad thing.
5. The community is the most important aspect of Danbooru, not the images—the images don't upload or tag themselves. It is clear to me from reading through the posts up to now that much of the community simply does not want AI art, for reasons both good and bad, while the people who do want AI art don't care that much. The community should not be put at risk.
6. There is still some value to AI art of Danbooru being focused on human-created art. Contrary to what many believe, generative models can learn fine from the outputs of themselves or other models, particularly if they have been heavily prompted or filtered by humans, and this is a standard technique for improving them; to the extent that the outputs are wrong, they can simply be treated as a 'style'; eg the way that large image models have learned 'Deep Dream' as a style. So it's not a fundamental problem for generative models if Danbooru contains a lot of AI art. However, it does confuse things in various ways, and if the quality is low, the future models may have to waste additional compute learning the 'style' of the glitches. So, since data is not really a bottleneck for anime models at this point (no one has yet trained a model which has overfit the Danbooru corpus, much less other image sources), adding AI images to Danbooru may wind up hurting on net rather than helping. So even from the AI researcher point of view, banning AI images may still be useful.
So overall, yeah, I just don't see much reason to be upset if Danbooru chooses to temp ban AI art. Looking at the long-term, I see a lot of good reasons for that, and if it is a mistake, it is a minor one. One might say that the ideal time for Danbooru to drop the ban will be when it becomes moot because no one can tell any longer whether an image was AI-touched, and all the artists are just using it as an ordinary tool, no more remarkable than the many specialized tricks Photoshop has supplied for eons.
Pixiv has put out its own statement regarding AI generated art: https://www.pixiv.net/info.php?id=8710
the tl;dr is that they'll allow it, but it will have it's own category, can be filtered out based on user preference, and have separate popularity rankings. Subject to change as things move forward. It still relies on AI art generation accounts self-reporting as such; although at the moment it's not like it's that hard to tell what's AI generated for the most part.