I choked on my drink when I saw the "restoration" of that old photo, was not expecting that

This feels like a modern-age water mark

i hate ai because a lot of people don't use it as a tool but to replace art effort

I feel bad for Greg Rutkowski. He's the most AI-emulated living artist. His original work is now vastly outnumbered by AI ones, and some searches almost exclusively give AI results when you look him up

"I don't need to explain why this will happen; you know I am right." Pure poetry and simply true.

I'm currently pushing my art skills to the next level and I don't think AI will ever match the sheer joy of making art and then continuing to improve

I work in a large game studio, and my boss trained a model using my colleague's artwork, then named the model after himself, since the art was supposedly made on company time. (He actually used his personal Artstation works, as well. He then fired anyone who was openly against AI (3 artists in total). Suffice to say, the results were bad, and the parent company fired that boss for unrelated reasons. But creativity hasn't recovered.

This coming from the guy that replied to my previous comment, telling me to "just use AI" when I brought up concerns.

About nightshade, the results have yet to be reproduced. People trained models using poisoned artworks and didn't notice any significant effects. So there's that...

The issue I see with this approach is that a painting, once poisoned and published cannot be "updated", where classifiers and models are getting better over time. Training data may be corrupted now because of these images, but I suspect technology will advance fast enough to be immune to this in no time.

AI generations and true art is starting to turn into the -Hacker- Cheater (sorry) vs Developer war. A developer blocks a way to cheat, a hacker finds a workaround, rinse and repeat.

My machine learning class taught us that in a generative adversarial network(GAN), any method of detecting an ai model can be used to train the same ai model in theory. GANs can also be used to avoid poisoning and can lead to an arms race of using the glaze itself to train the ai model, and glaze learning to become better at poisoning. This is really only a temporary solution until the training model catches up using this technique.

I imagine someone will just train a denoiser for images generated with nightshade and glaze to get around this.

I'm just a hobbyist and not a professional artist so this doesn't affect me any really but I can emphasise with people who do make a living from their art. I think AI art is fun to mess around with, or possibly as a tool for reference, but I do hate seeing AI art flood art sites. I want to see something actually drawn by a human. Even with some art sites having settings to supress AI art not everyone properly tags their art as AI. Not to mention people selling AI art or running patreons with their AI art etc. Most of the ones I've seen are just low effort and the basic stuff you get from typing prompts. I know cause that's the sort of stuff I got quickly messing around with some AI image generators. You can pretty much tell AI art instantly depending on the style. Anime art in particular seems to have a distinct AI style.

I don't like the idea of having to make all images of art lower quality. And AI is definitely gonna just learn and adjust to this

it's a nice thought, and I understand why artists are excited by these efforts, but the fact is 1) no one has been able to reproduce this poisoning effect yet, who knows if it even works on say, SDXL-turbo, or an already finetuned model. 2) it would be trivial to download 10,000 images from laion, run the poisoning on them, and train a model to convert poisened to non-poisened images. Then in training, you simply add a new step to "de-poison" the image. Bypassing this isn't a matter of if it can be done, or even when it'll be done, but if anyone bothers in the first place.

That uncle restoration caught me off guard, can't even lie

congrats for winning the best content creator award philip!

Sadly i think the battle is already lost How many millions of old images are there on the internet from inactive or unaware artists. While it will be harder for a ai to learn new styles in the future, the classical realistic styles of 1 million google results of dogs before 2020 remain.

Jumpscare at 6:47

Thanks philip. I'm having a panick attack because of 6:47

Some report it takes from 20 minutes to 12 hours to process an image with that (tool is based on AI model). Also "Use deepbooru for caption" and "Use BLIP for caption" both still tagged nightshaded images fine when I tried, so it does not seem to help against auto tagging tools. AI training specifically is a process, and you need to both tag images and cut / scale them to specific size squares before the training, and I feel this step alone tends be enough to limit the effectivenes of any pixel alteration based method.

It's so bizarre how many techbros will just approve of unprecedented mass copyright infringement because it makes the slop get made a little quicker. People talk about dystopian futures due to the government/technology etc. but the people who will bring forth that dystopia are americans, willingly.

Every time I see AI Artworks on Social media, I try to mute those "artists". I have nothing against them, but I'm not sure if their AI "art" is legit or not

Watching AI artbros moaning about how glazing and nightshading is "dRiNKinG pOiSon hOpiNg tO hUrT oTheRs" is priceless. Oh, and they are starting to demand their prompts be protected by copyright laws too, the audacity of those parasites is just stunning to watch.

it might stop AI from stealing your particular art, but it won't stop AI from lowering your wages

Great timing. Saw stuff about nightshade but didn't find a good video on it. Now I have one

I geniunely appreciate you making this video cuz I feel my art being used by someone for AI is unavoidable at some point during my art journey, this is a good video I'm saving and it is definitely useful for me when I post my art online but keeping the clear version and the original files myself. I really didn't expect that this is something I would learn from you on this sort of matter xd

Thank you philip. As an artist and a long time follower of yours I really appreciate you putting light on this :-) I look up to you and really admire everything you do^^

I am not an artist but I enjoy the the stuff PEOPLE create. These days, if not completely obvious, I just don't trust artists that have no works posted before 2020. If there are more people thinking like me, it means that becoming an artist now has become much harder. Also, art that looks the best gets hit the most.

The problem with both these options is that they fundamentally alter the picture and introduce noisy artifacts. Passing a picture into some edge-refining tool would most likely break this. Obviously it blocks data scrapers trying to automate the process, but anyone with enough time to actually want to copy a specific artist would probably be able to do so without too much hassle.

Great video, learned something useful for an essay I'm currently procrastinating writing. I'm sure It'll come in handy soon enough, Thanks !

unfortunately, if glaze gets popular, a de-glaze AI will probably be developed to restore the original for training AI. I don't think it would even be that difficult for them.

I loved the Dr.Suess-esque image "cockification" at the end - great video!

Midjourney devs got caught discussing laundering and creating a database of artists to train off of which is now evidence in a current lawsuit, i dunno why anyone thinks it's fine to use copyrighted work without permission in commercial products? it's not like copyright law doesn't apply to these companies because of new software?

Glaze + NS would probably fill the same niche as a watermark. A paying customer would probably want the original unaltered copy, but those watermarks aren't _for_ paying customers.

On a serious discourse I really doubt this technique will have any intended effect any effect on large model training, these techniques work by using a small proxy CLIP model, and the hope is that if it works on this model it will also confuse larger, trained on a different dataset model with potentially a different architecture, while not introducing too many artifacts. Nightshade promises was also to confuse auto-labers, but now that it's out and even with the strongest setting I can't manage to fool GPT-4V or even open source and much smaller model like LLaVa-1.5, as some other paper has shown the bigger the image encoder became the more align is to human vision. In the end I think some people will spend tons of compute to create images with adversarial noise that will help the models learn more robustly and not take shortcuts like the proxy model did (the reason why it was vulnerable), as the images with the adversarial noise or not will be labeled correctly the model will actually learn that the patterns that deceived the proxy model are not what a human would think a cow is for example, and the model should be able to align more to what a human would expect.

It is true that AI isn't intelligent. It isn't creative, it's still a deterministic program, it can't be spontaneous, it can't create new things. AI is a heavy deceptive marketing tool, AI can be very useful, but current it will not be as good as humans, it will never be as long as it keeps being deterministic. AI IS DETERMINISTIC. We MUST protect people that actually create new things.

As a temporary measure there should be some kind of custom small watermark or something like if you include this string of numbers or text in the uploaded file the model wont take it

When I first saw glaze, as shown in your video, I actually thought it was AI generated artwork, before thinking instead that it was JPEG artifacts. Not sure what that means for the tool if the first thought that someone has when they see glaze is that it's AI artifacts.

most of the art i see are already reposted so many times that it's a crunchy pixel mess so this filter wouldn't really change how i see the work anyways.

God dammit why do I think your humor is fantastic- i love that you embrace it without going too far off the rails

I'm so happy that this exists. I'm so depressed that it has to.

the weird thing about the glaze thing is that it distorts it in a way that could be confused for ai generated

re: would they notice? - ive seen larger errors in ai generation go unnoticed like strange architecture, the amount of teeth in a face, finger shapes on hands or the way foliage is too chaotic even for nature. as an artist myself i feel that if a work of art has those defects after being run through glade and/or nightshade it would be indistinguishable from being an aspect in the style of which the artist renders their work, though it may be more obvious if the art isnt a painting.

This is why making irl drawings and paintings can save you.

Also would it not make more sense to nightshade the image after glaze, since glaze is much more destructive in the final result?

I like that you are using AI art for the video thumbnails. The artist has two left hands.

Its really shitty that companies are having a "ask for forgiveness not permission" model for AI stuff, you should ALWAYS get permission first.

Nightshade and glaze the good AI

You can also double enchant your image so half of the image have glaze and the other half is nightshade. So plus +2 enchant 😂

I'm not sure to have understand it correctly, but it looks like it only work on datasets that use CLIP to transcribe the text, since it depend on making few changes to confuse it, but any new CLIP like model would be immune to this specific technique, maybe in the future the dataset creating will create some kind of consensus between various models to create the description of the image, that's seams like the most effective strategy.

I GOT SO SCARED WHEN I HEARD CABOOSING AGAIN. PLEASE STOP

Genie's out of the bottle and it's not going back in regardless of what anyone wants. The conversation should be about how artists can get compensated for their work going forward, but trying to "stop" this ain't ever going to happen. It literally _can't_ happen. That's not how disruptive inventions like this work. _Someone_ will use it and that's that. I mean, remember Metallica trying to stop Napster? Doesn't work.

ever since the video on the song that apparently everyone hates (which by the way I like and didnt think was weird) I feel like you play it at higher volumes now. Maybe it's just placebo, until you make it really obvious in some csgo video where it starts at like 400% volume then fades out. I appreciate this gag lol

Haaa videos for youtube usually cut off at the end to not lose so much viewer retention (i think?). This video actually used that technique in a perfectly fitting way.

Imagine if Nightshade and Glaze would be implemented in to art programs like Krita or Blender

It's interesting how you were the one of the first people to be on broad about AI-content in general and now this, glad to see you capable of thinking for yourself.

The time has come that physical artworks have become more valuable again

Thank you for covering this; the war has just begun -Nightshade is a fantastic first offensive -and if any of you who sympathize are in the States, please consider donating any amount to Concept Art Assoc.'s fundme for protecting artists against AI. They are making wonderful legal efforts

The human urge to spew mass produced forgettable garbage at the lowest effort possible.

this sort of controversy gives me very similar vibes to react content, where ai training seems to be taking content without permission and then only taken it out after request, which is what a lot of reactors do. i wonder about that for you philip as im sure you mightve been a victim of the reaction content, whats your stance between the two? you feel the same for both? or do you think its different for original youtube content being reacted to vs original artwork being trained on?

As an hobby artist this is quite the enjoyable development, looks like we are getting more ammo for our side of the war after the corporations caught us off-guard at the start

"Wild West era for AI" is the best way to put it

It's a nice thought, but unfortunately I really don't think it's gonna be enough. AI models keep being updated and eventually an anti-poison feature will simply part of the software.

I think glazing will create demand for prints. If a user wants to commission an unglazed image, they can pay extra to get a physical copy. People that want a beautiful SFW artwork will likely pony up the dough. On the other hand, the porn artist will still be raking in with the glazed images since you don't need the pic to be high rez to wank to it. But really, there should be laws to protect art from being used as training data. I feel like the best solution for both parties is if the artist is given the right to sell rights to use their art in training data, AI could instead become an impetus for artist to make money instead of being put out of work.

This is a good idea. How do I apply glaze and nightshade? (Especially nightshade, that one sounds absolutely hilarious)

Ok. So what about all the art I have published already that I can't replace the core files of without compromising the metadata of my entire art profile with false upload dates? Do I just change my entire style in a massive pivot and leave all the things I've made up until now to be trained on and replicated? I get how this might help artists that are looking to get into having a web presence but artists that already have published art are kinda screwed here regardless unless each hosting website offers core file replacement options. I see great benefit in keeping a low profile for now, not inciting enough attention to incur copycats until this wild west situation calms down and there's proper means to go about things. How about an art profile that requires a state ID to sign up? Artists could post their artwork to their confirmed official outlet of internet publishing and viewers can just cross reference whether or not an outside work has appeared on the official page of the artist... a page that's hopefully locked behind a good password and two factor authentication. This solution is going to incur a lot of stress upon people that don't have the living situations required to sign up for those kinds of accounts though, so it's really just a band-aid. Operating systems could add features for automatic cross referencing photos with official artist pages, flagging photos that haven't appeared on official accounts at any capacity yet. A better system for ensuring authentic official artist accounts can get workshopped but this is the best I've got in the meantime. This way we can have a bit of our cake and eat it too. We could reupload pieces we've posted elsewhere before to our official pages while leaving the original uploads with their metadata intact. With something like that safeguarding artists, in theory, and I mean this is just a theory, it might not even matter how good A.I. art gets at replicating the works of artists. Every style will, in a sense, be a bit like how music genres are. Pastiches. And if you want to hear an artist's official work rather than a random selection from a genre they happen to partake in at some capacity you search up their name and play what shows up on their official profile. Like that, but with visual art. It's not just looking for the pastiche but one of the OG's takes on the pastiche they literally made. Meanwhile, maybe some kind of image cross reference A.I. that detects styles and finds the earliest upload of anything in that style with upload metadata could help. It'd be a lot more broad, it'd have a lot of false positives, but it could do some good in giving some kind of pipeline for people that enjoy the style of an A.I. work wanting to engage with the artist that originated that style. There's hope. Art will not die, it will only get different. We may even reach a point where these generation tools won't be so widely available that this kind of protection has to be as commonplace as every instance of a common person. This is the wild west right now, and all of this is going to be looked at like a highly controversial art movement on both sides of this.

Glad you dedicate a video about this side of the conversation as well. Must say I was a bit annoyed with the general publics praise of generated imagry that came at the cost of using copyrighted artwork, while using it before for commercial purposes wouldn't be dared. But developments like this make me hopeful that, if not the usage of copyrighted artwork in models will be unusable, it'd atleast become banned for commercial usage.

"Shit on your food to stop other from eating!"

I feel like we are currently in the "Gilded Age" of AI, where anything goes and so we have some crazy huge, impressive models that can be trained fairly easily. However, we WILL have to put some restrictions on AI, like what happened in the Gilded Age. It probably won't happen until serious damage has been done, though, because that's the way policymaking tends to go.

I guess the glazed versions are like mona lisa replicas and the normal ones are the actual one itself or something

Only publish poisoned images and only send to customers a picture with an imperceptible watermark (I'm sure it can be done) so if it leaks you know who leaked it and sue

Ai is already at the point where it can use generated images to improve, so there is all this effort being put into an area that is already being pulled away from. and as you mentioned the individual style copying is going to be done by individuals, and they will just pick images that aren't poisoned. and unless the artist never shares/sells an image, it can be acquired and used to train a personal model. I have looked at the paper about poisoning the training data, and I feel it would only sort of work if the ai researchers/developers stopped working on their projects and sabotaged their own work. Like part of improving the technology would be making sure the quality of the incoming data is good, and it would detect that it is either a glazed image and categorize it as too low quality to use, or if it is label poisoned like with the dog/cat thing, then it would see that the label doesn't match what it has seen in the past and also throw it out. and I guess you have succeeded in having an image not be used in an ai model, but you have also either reduced the image quality for all the humans you share it with or ruined the accessibility of the image by having the wrong description of the image. I support artists and enjoy making art, I just don't like spreading false hope, or information that doesn't seem to have much to back it up.

amazing bro keep up the content! on all channels !!

The fact that this is quite literally an arms race for mathematicians and engineers almost makes me excited for all of this

This is truly a tough problem. Artists are absolutely justified in bringing upset about every going on with AI and training on their work for free. However, I fear that the cock’s out of the bag, and there’s no getting it back in. Wait… cat?

There are already ways to "de-glaze" images

Objectively, the only thing hollie not giving permission could achieve is making her look a lil grumpy. High-res images still exist on the internet regardless

" instead of coks we will get coks "

So, for those wondering, here's how the current process works. Model trains, makes image, different model looks at image and says if it looks off and points out flaws, model then trains on this feedback. The "poison" system is probably just exploiting subliminal noise patterns which is literally just a bug. The entire issue they are "working with" (exploiting) will likely be completely fixed/solved in the next few months at the longest.

Dont know if going scorched earth is that great of an idea. Its going to damage so many beautiful pieces.

Overall interesting food for thought, thanks for sharing your views!

Taking a picture of the screen, checkmate Holly

>how to protect your art from lower quality AI replications with no soul? >by lowering the quality of your artwork by adding artifacts to it thus hurting its soul idk if this is a serious suggestion or satire. I hope its satire.

I believe the entire concept of intellectual property needs to be re-examined thoroughly. No one wants the artists to starve, but at the same time, if we go really hard with AI licensing laws, only gigacorporations will be able to buy enough training data to create AIs. It will end the open-source grass roots level AI development right then and there. I believe that to be the more dystopic alternative.

sir u never stop amazes me for your creative content

I mean it's not like poisoned images will remain poisoned forever, modells will adapt, it's just like DRM anywhere else: and endless cat and mouse play

Skimming the comments I have to say people don't understand copyright. It does not give absolute control to the rights holder, as much as many of them would like. Lines will have to be drawn, and that will take time for the legal system to work through it.

Given that artifacts are visible, detecting it might be possible. As I understand training happens preferably on high quality images. I also strongly doubt the effect could survive just cropping the image by like 10% and resaving as jpeg. But that would all have to be tested.

As someone who just recently started getting into creating art, AI does kind of discourage me at times.

1:23 Remember guys, that is future. Think trice when you take selfie.

I don't really care about the results of these arguments, but I do love hearing them. It really makes you feel like you're in the future. I wonder, if AI does kill professional human art careers, if that will force humans to do art for the joy of it alone, which is probably the healthiest reason to do art in the first place. Time will tell

So glad the default state is "your art will be stolen by default without consent. Click here to stop that."

When you put two things on screen to compare them, can you keep them up for a little longer so I don't have to pause it at just the right second

The ending had me wheezing, and i know you're right!

if you can add protection, you can use the same algorithm to detect those images

this actually makes some SD models better at rendering. its hilarious

{Literally 2 hours later Anti-Nugtshade Torch tool was added to most AI midel generators}

A simple uspcale of any image using SCUNET before training would be enought to completely break this "security" and now you have artist using worse images to show off their work. That is not what I call a way to protect artist's work. We just need regulation and ways to detect AI art in general, especially in pro work.