Official twitter for UChicago Glaze/Webglaze/Nightshade to protect artists against invasive AI. News/updates. Mainly
@ravenben
@shawnshan_
2.2M/250K downloads!
Today is the day. Nightshade v1.0 is ready. Performance tuning is done, UI fixes are done.
You can download Nightshade v1.0 from
Please read the what-is page and also the User's Guide on how to run Nightshade. It is a bit more involved than Glaze
By now, I'm guessing most have already seen the news on our new project, Nightshade. Lots of artists sharing it, but here's the article from MIT Technology Review (thank you to the wonderful
@Melissahei
), and a thread explaining its goals and design.
AI bros pushing hard on misinformation campaigns that claiming Nightshade is illegal or its users can be held liable for harm to ai models. I can't get over the entitlement needed to think everyone on the Internet must make their content suitable for your profit taking scheme.
Well that didn't take long...
Glaze 2.0 is here. Improved robustness for all art styles but particularly flat color styles. Addressed patch artifacts. Improved efficiency (~50% speedup) and support for Mac GPUs. Mac M1/2/3 support is here, Mac Intel coming this week.
Glaze news:
We're getting ready to do a major update to Glaze. Glaze 2.0 makes fundamental changes to the core glaze algorithm to improve its robustness on all models, increases protection strength for anime/flat color styles, and even runs a bit faster than Glaze 1.
Artists, calling your attention to a new tool. Mist works internally similar to glaze, but demonstrated stronger results against Img2Img. We believe bec it runs many more iterations than Glaze's maximum rendering setting. Cool results & glaze will add these settings n next update
Quick thread on common questions.
1. no, NS does not scrape/copy/train on your art. It works perfectly fine in offline mode. Install, turn off wifi, then run. After dedicating 14+ months to this, folks who accuse us of trying to steal art will be blocked without a response.
Quick notes
1. probably not best idea to announce an image as poison, kinda defeats the purpose
2. a shaded image is not glazed. If you want protection from mimicry, do not post your art with only nightshade.
3. NS is memory sensitive. If u have big apps running, close them first
Hi all!
We've been busy on Glaze related projects. Happy to announce that webglaze, the Glaze web service is now ready for new users. If you already have a GPU and it works well, then please keep using that. If you have a mac, or no GPU or no computer even, webglaze can help. 🧵
Another big day. The big big Glaze 1.0 update is out. New core optimization function makes Glaze MUCH more subtle. Now works well for most flat-color styles, comics, anime, cartoons. Nearly invisible on most textured art images. Also brand new website!
Folks. This version does not include glaze. We will work on an integrated version of glaze+ns. For now if you want benefits of both, you should shade first, glaze last. But note that artifacts might be more visible. The integrated version will fix this issue.
Glaze news:
We're getting ready to do a major update to Glaze. Glaze 2.0 makes fundamental changes to the core glaze algorithm to improve its robustness on all models, increases protection strength for anime/flat color styles, and even runs a bit faster than Glaze 1.
It's such a privilege to be able to work with some of the most skilled and talented artists out there today. Greg and many others who not only produce breathtaking artwork, but are so generous in their advocacy to protect other artists in the field.
It’s been a while since the Nightshade premiere, but I finally had a free while to apply it to one of my recent paintings and I strongly recommend to you to do the same 😊 Link to the Nightshade:
#nightshade
#noai
Hi all! Nightshade tests continue. We would also like to ask for artists' help on a new research study that evaluates the ability of artists and software tools to reliably separate generative AI images and art by human artists. Please help us on a survey:
ok ok, 1 last tweet I promise.
I realized the most surprising result was not included in the MIT TR article. You can read the details in the paper (fig17), and I will just leave the figure here. FIN/
The reception to Nightshade is overwhelming. Thank you to all the warm comments and kind messages. Among all this, Glaze is still a thing, and this bit of news came out today. Glaze was chosen by TIME Magazine editors as special mention for Best Inventions
Highly recommend you read the User's Guide to understand how to use the poison tag.
NS is very similar to Glaze in general tools used. So the usual robustness applies. Cropping, resampling, compression etc etc have minimal impact on NS poison.
Have fun!
This is... upsetting
From what I know, Univ of Chicago receives requests from AI companies for our students/faculty data all the time, and takes an active role in preventing our data from being scraped / trained. But a school actively offering student/faculty data for $ 😠
As more artists download glaze or signup for webglaze, a reminder that Glaze is a constant target for attacks seeking to weaken/bypass/remove its effects. We will let artists know if anything succeeds along those lines. We can/will evolve as necessary to adapt to countermeasures
The demand for nightshade has been off the charts. Global Requests have actually saturated our campus network link. (No dos attack, just legit downloads). We are adding two fast mirror links for nightshade binaries.
Win:
Mac:
We are happy to collaborate with the great team at Cara to help protect the works of human creatives, on a platform for artists, built by artists. CaraGlaze has all the features of WebGlaze, and will get future updates as they become available (incl. Nightshade).
Hi artists! Want to protect your work from style mimicry by generative AI? I'm happy to share that the long-awaited collaboration between
@TheGlazeProject
and
@Cara_HQ
is finally here!! 🥳✨
A quick thread on Cara Glaze and Glazed art in action! 👇 1/
9. I have talked to plenty of lawyers, none of them have any concerns about legality of me creating NS, or anyone using NS to protect their own art. Putting hot sauce in your lunch that is clearly labeled "DO NOT EAT/SCRAPE" does not make you liable when someone steals your food
Quick announcement. The Nightshade paper has been peer reviewed and officially accepted at IEEE Security & Privacy, one of the big 3 computer security conferences. May 2024. We will update the arxiv paper w/ the camera-ready when completed.
A few updates:
1. WebGlaze is now updated to Glaze 2.0
2. We are hiring a student to handle webglaze account requests. We will work thru the backlog of weeks of requests. Thanks for your patience
3. Glaze is NOT open source. There is a fake Glaze repo on github. They're not us.
Hi all.
Quick update on Nightshade. We have a beta app, but it is still using too much GPU memory, and the user interface is still messy. I'm sending it to a few people for closed beta testing. Hopefully we'll get this issues ironed out soon.
Thanks for your patience :)
Hi all. In addition to Glaze2.0, we are working on a project extending Glaze-like protection to short videos and animations. (Our prior user study led to a paper to be presented at CCS 2024).
We need volunteers to look & rate the quality of short video clips. 5-10mins max.
What is Nightshade? It's a tool that performs a data poisoning attack against generative AI image models. Poisoning is not new. Poisoning genAI models at scale is new. You can read the MIT TR article for the high level story. For details, here's the paper:
Just wanted to say thank you to all the wonderful artists who have helped make the Glaze project possible. We are honored to announced that Glaze received a Distinguished Paper Award at USENIX Security Symposium, and the 2023 USENIX Internet Defense Prize.
Note to anime/manga artist friends. Glaze effects will show up easily on clean-color styles. So please try very low intensity values if you can. start with 0 or 3. If you get errors, try several times before you give up on a specific intensity value. Anything that passes is ok!
This is so important. Artists are the canary in coal mine, because images are easier to target, & artists have less labor organization/resources/lawyers. But creativity is aspirational and a core part of being human. Ask anyone you know who creates for a living about genAI.
In this week's newsletter, I argue that artists are the canary in the coal mine. Their fight belongs to anyone who has ever posted anything they care about online. Tools like Nightshade could be a first step in tipping the power balance back to us.
Why Nightshade? Because power asymmetry between AI companies and content owners is ridiculous. If you're a movie studio, gaming company, art gallery, or indep artist, the only thing you can do to avoid being sucked into a model is 1) opt-out lists, and 2) do-not-scrape directives
Hi all. We learned today there is a system called Sieve which is a commission platform for artists. For some reason, they choose to call their protection process "glaze." Please note they are not affiliated with our lab, UChicago, our group or the Glaze/Nightshade projects.
News for our artist friends in Japan. We have a complete Japanese translation of the Glaze webpage:
Special thanks to the fantastic Mayumi Nakamura Birt (
@tinynakamura
), who found time out of her busy schedule to help us. tysm!
Hi artists!! We hope folks are finding Glaze easy to use. But if you have any issues / questions, please reply to this tweet and we'll follow up. You can also DM this account (DMs open) if you want to send messages privately. Thx to
@ZakugaMignon
for the suggestion!
#AskGlaze
That's the same attitude as the office lunch thief who is upset that you didn't make your sandwich the way they liked it. Too much hotsauce on the lunch I stole from you, pay my doctor bills!! Why couldn't you trim the edge off the bread my way?
Hi all. An update on planning on increasing availability of Glaze to more artists, especially those with less or no compute power. We have an amazon grant for AWS credits, and are considering hosting a "web service" for FREE to artists. No art is stored, all deleted afterwards.
Important to note that image generation is THE most expensive (in environmental harm) of all generative AI tasks, which in turn are much worse than traditional AI (e.g. discriminatory classifiers that don't have ethical questions and piles of class action lawsuits).
The energy and carbon costs of deploying AI models have largely been unknown.. until now! 🌏🚀
With
@strubell
and
@YJernite
, We tested 88 models on 30 datasets from 10 different tasks from different modalities and found some pretty cool stuff!
A thread 🧵:
Update on our study on the detectability of generative AI images. Thank you again to the 4000+ artists who responded to our survey. The study capturing your inputs and efforts was peer reviewed and accepted for publication at ACM CCS (3rd of the big3 in security), October 2024.
Quick update. Webglaze is temporarily done because of a cooling issue at UChicago data center. Our tech staff is working on it but right now most of our servers are offline to avoid overheating. Hope to get them back soon. Stay tuned.
Quick update. A big thank you to all who participated in the survey we posted last night. We quickly received nearly 2000 responses, significantly higher than our target. Mission accomplished. We opted to turn it off to limit exposure of real art to potential scrapers. thx again!
People making up stuff about Glaze is not new, but claiming Glaze is genAI is definitely new😂. If anyone is wondering, Glaze includes licenses for StableDiff, because we analyze it to identify its weaknesses to guide Glaze's disruptive changes. Glaze has no genAI & no training.
@pehdrigues
@TheGlazeProject
no it’s not.
They scraped the internet to make Gkaze too.
By using Glaze you are give them permission to use your images. The same way you give Instagram the right to use your images too.
None of these mechanisms are enforceable, or even verifiable. Companies have shown that they can disregard opt-outs without a thought. But even if they agreed but acted otherwise, no one can verify or prove it (at least not today). These tools are toothless.
@GenelJumalon
@palomamcclain
Just to put a simple label on this. The former requires ONLY discriminatory classification, ie “traditional” AI, and the latter is generative AI. People need to understand that nearly all the major selling points for how AI can do good is traditional AI, and NOT generative AI
Update on this. Cooling system came back late last night. All servers were up this morning, and I just turned on the web server/webglaze service.
File servers hosting downloads are also up and restored to normal. Thanks all!
Quick update. Webglaze is temporarily done because of a cooling issue at UChicago data center. Our tech staff is working on it but right now most of our servers are offline to avoid overheating. Hope to get them back soon. Stay tuned.
Announcement. We finally released the v1.1 update to the Glaze app. This now updates the app to match WebGlaze with our newest, most robust Glazing algorithm. Download Glaze App v1.1 for Windows/Mac, here:
hey all...
Heads up that we might take webglaze offline for a bit (and folks should pause their Glaze app too). We've been running tests on our "self-test" that determines when glaze is producing strong enough effects. The findings are more inconsistent than we expected.
Indeed. Glaze has over 658K downloads in 8 weeks, about 11,000 downloads per day. We're working on more improvements to come soon, first of several to improve glaze on comics styles, and more tweaks on preserving metadata.
ARTISTS, great news from
@ravenben
from
@TheGlazeProject
. Since March 15th there have been 658,298 downloads globally!!! which is about 10k downloads a day!! Keep protecting your art!
Hey all!
Been a crazy week, but we are basically done with significant changes to Glaze internals, following our week long investigation into our self-test. WebGlaze coming back online tonight. Have to get sleep & will be giving a talk at Columbia Univ CS tomorrow.
Hi all. Sorry for the delay on Nightshade. Students mostly off for the holidays, and some are now trying to recover from covid contracted during holiday travel. So we have not yet started on changes to address feedback from the closed Beta. Reminder for all to mask & vaccinate.
For artists interested in using WebGlaze to protect their art from generative AI models, please follow (to the best of your ability) the instructions here:
This will greatly simplify the process and help save us time. Thank you!
No artist is able to force training data into a model. Only model trainers can voluntarily *take* your images (shaded or not), against your wishes, to train their models for profit. This is not a virus, it does not infect/affect any model against their wishes.
😝 PS: folks, Nightshade adds a small amount of poison to each image. Its effects only take over a model when it ingests a lot of shaded images on the same prompt/concept. For those claiming NS is ineffective after training on a single image, please go read the paper first.
2. Yes we know many artists run on mobile devices. No we will not make android/iOS apps. Yes we WILL integrate NS into WebGlaze when we finish our tests.
3. NS has same robustness properties as Glaze. Cropping, resizing, screenshots, pixel smoothing all have minimal impact on NS.
So now everyone knows why we've been so quiet for the last few weeks. With the webglaze update to v1.11 and Nightshade, that's why we're 1 month behind on Webglaze invite requests on twitter (sorry, we're working on it!)
And just to be clear. Yes, I have talked to a number of lawyers, both private sector and public sector, including the UChicago university high counsel. Also had the pleasure of giving an invited guest lecture at the law school in the Copyright class. That was a great experience!
We expect artists and content owners to make use of every mechanism to protect their IP, e.g. optouts and robots.txt. For companies that actually respect these, Nightshade will likely have minimal or zero impact on them. Nightshade is for the other AI model trainers who do not.
The Nightshade paper is public. Any one is free to read it, understand it, and do followup research. It is a technique, and can be used for applications in different domains. We hope it will be used ethically to disincentivize unauthorized data scraping, not for malicious attacks
I work regularly w/ a number of lawyers on legal matters, and also with regulatory agencies whose main teams are mostly made up of lawyers. And not a single one of them is worried for me or the artists I work with. So I really appreciate the unsolicited concern, but we're good!
The Concept Art Association is fundraising to gather resources for numerous advocacy efforts on behalf of artists/creatives everywhere. Reminder that without CAA's efforts (like this townhall), Glaze would literally not exist. (You can see birth of Glaze 1hr 26 into this meeting)
3/ Back in 2022 the CAA was one of the first artists orgs to realize how detrimental & exploitative GenAI was to its community. They quickly mobilized via townhalls like this one
(fun fact you can see the early beginnings of
@TheGlazeProject
at the end!)
We are currently considering how to build/release a potential Nightshade tool. It might be integrated into Glaze/Webglaze as an optional enhancement. We might also, time willing, release a reference implementation as open source. Stay tuned for updates, hopefully soon.
FIN/
More importantly, only large model trainers have any incentive to consider opt-outs or respect robots.txt. Smaller companies, individuals, anyone without a reputation or legal liability, can scrape your data without regard.
ok, 1 more tweet.
Super important to note this is a big (and growing) team effort at
@UChicagoCS
, with absolutely amazing and critical help from so many amazing artists. You know who you are. Those of you who took that first Glaze survey might remember the last few questions ;-)
@InfernoOmni
Really? what do you think an AI specialist requires in terms of training? Remember the loud voices proclaiming prompt engineers as the next 300k/year career? Yeah, me neither. GenAI is a hype machine leaving only destruction in its wake. There is no job security in AI.
Hi all.
Apologies for the LOOOOOOONG delays for your webglaze account requests. I've been working on research deadlines and have not been able to put time in to invites. Trying to hire a student to handle these full time. Hopefully this week or early next!
Hi everyone. An update. We're finishing up some tests. but they're confirming what we've suspected, that CLIP score we've been relying on has been decreasing in usefulness... it's making some arbitrary decisions on protection strength that do not match end to end eyeball tests.
Yet another wave of misinformation on how pixel cleaners can remove Glaze. More info: 1) the original adverseCleaner GitHub page already admits this does not work. 2) we mention this on our FAQ , and 3) we tested it & included results in our paper.
PPS: thank you to all the wonderful artists whose creativity inspire us daily in our work. Also
@katriaraden
@zemotion
@zipavika
for amazing art. Artists represent the best of humanity and why we do what we do. Final reminder that I’m the social media manager who can’t code and…
It's worth noting that the MIT TR article includes a result that represents the "dumb" version of the attack. The optimized version that is Nightshade requires roughly 5X fewer samples to achieve the same result.
@katriaraden
@TheGlazeProject
It would take thousands of images to even affect a base model (which is impressive; a fraction of the billions of stolen images could ruin a dataset.)
If enough people use this it could work. I doubt it, but its a nice thought :) Hope I'm wrong, this could be a great tool.
4. No time to maintain Linux version. Sorry. We will explore cmd-line version later.
5. For samples of before/after, we will put more up on nightshade website in a bit.
6. Sorry downloads are slow. A lot of demand is actually saturating our network. Never seen that before.
Helpful tip from
@aimeecozza
. To minimize finegrain artifacts, send PNG to Glaze and get PNG back to minimize JPG artifacts. Almost like processing photos in RAW.
Also: If your style is v sensitive to glaze, rerun a few times & pick best one. Glaze has random component each run
13. Finally, I understand the urge to lash out at model trainers. But please be responsible. Sign up for opt-out lists and use do-not-scrape directives. Make sure that if a model is poisoned, it is because they went against your explicit wishes to not be trained.
Hi all. Big update thread.
We put webGlaze into maintenance mode a few days ago to test/rethink our self-test mechanism. The result is an update (v1.11) that removed the self-test in favor of a precomputed intensity level for each art style.
Hey all. More Nightshade news!!
We (by that I really mean
@shawnshan_
) did some great work and implemented a workaround for PyTorch issues on Mac GPUs.
New download is now available for Macs. Nightshade 1.0.1 supports GPUs on Macs with M1/M2/M3 chips.
7. Unfortunately NS still relies on PyTorch, so Nvidia GTX 1660/1650/1550 still do not work. Sorry. We will eventually fix this in time.
8. If you cannot wait for glaze/NS combo (forthcoming), then we recommend shade first, then Glaze to prioritize Glaze protection over poison
11. Sorry we have not yet made a package for Intel Macs. We will work on that shortly. Should not be a long wait.
12. Is it worth it to glaze/NS images already scraped by models? yes, models are always looking to scrape more images. They often have trouble identifying duplicates.
Sigh. Apologies. But glaze app v1.1 has a bug. It glazed too hard, at too high an intensity. We fixed the bug for Mac version (updated online to v1.1.1) but the windows version will get updated later today/Sunday). Please download the v1.1.1 version later today.
10. In terms of image quality, nightshade uses same perturbation limits as Glaze (and even lower intensity values are allowable). So visual artifacts are at most equivalent to Glaze, and in most cases, more subtle/harder to see than Glaze effects.
Hi all. Realized today we have not updated folks on Glaze things in a while. There's so much happening on Glaze and related topics. Some are definite, and many are still evolving. So here's a thread. 1/n
hey all. Thx to all the artists who responded to our research survey. Instead of sending out thousands of emails, it's easier if folks come and fetch their survey results from our server. Just type in the same email address you left us here:
Thx again all!
One final update then sleep. We've been a little behind on documentation. But finally put up a webpage explaining the design of WebGlaze, invites, using Glaze, common questions and general info.
It's here:
This seems to be the most recent and most complete article on the increasing phenomenon of unauthorized hijacking of voices. Another creative industry under siege by generative AI. We will study, learn, and see if we can help.
new: Voice actors are getting replaced by AI. These often faceless pros narrate audio books + commercials. Now, companies can buy clones of their voice, use AI tools to make them say anything + save thousands. The samples aren't perfect, but many may quit.
ahh yes, thanks for several of you who reminded me of the DMCA (1201) circumvention clause. This *is* actually a thing that I have discussed with lawyer friends. Here for your perusal.
And to be clear, no mention of AI in list of statutory exceptions.
Following the introduction of CaraGlaze, some questioned the decision to integrate with Cara. Isn't it safer to never integrate with anything? Sure. Adding any new stakeholders always increases potential risk, even w/ end-to-end measures like encryption, e.g. side channels. 1/n
Thanks to Akane for this cool and informative tutorial! Super clear and easy to understand!
I didn't even know it was happening until an artist forwarded an earlier tweet.
🔵Easy GLAZE USAGE TUTORIAL and some TIPS to PROTECT your ✨ ART ✨🟣
You gave me many likes when I asked if you wanted a tutorial on this, so here it is!!
But not explained by me, but explained by Lynna, my OC, (who suddenly knows a lot about these things.) ✨💙
Whew. a few days of non-stop invite generation for WebGlaze, and we're over 700 users. At least pace is slowing down a bit. Thanks to our artist friends for helping scan so many requests. Thankfully a number of folks had Cara profiles. Fast ID for human artists!
We will release an updated app and also update WebGlaze at roughly the same time. ETA is soon, certainly by end of the month and quite likely in the next week or so. If you're planning to glaze a lot of art, I would recommend waiting for 2.0.
In the last year, I've had the absolute privilege of working alongside amazingly skilled artists, who also happen to be exceptional human beings. I do not presume to speak for artists. But if there's anyone I trust to understand the issues & advocate for artists, it's
@kortizart
@matdryhurst
@neilturkewitz
You claim to speak and act for artists, yet when allies & artists continue to state you’re missing pivotal information, how your dismissive stance on copyrights and insistence of opt out is damaging, you then mute, hide and ignore? You’re not the representatives you think you are
Ok, the wait was not long.
has been updated with the Windows GPU version. The download is bigger (sorry), because ginormous PyTorch GPU libraries. But so far the speedup seems ridiculous.
Have fun. Let us know w/ comments here how well it works for you.
Super quick note.
We'll be making a windows GPU version for glaze very soon. Initial tests show it's a ridiculous speedup. So if you're grinding through a backlog of images to glaze, might be worth waiting a day or so. It's a bigger download but a LOT faster.
Kinda hilarious: Brian Cheung from
@NBCNews
visits our lab in person, and we use Nightshade to turn him into a cat (with his consent lol). Also a great interview with the amazing
@kortizart
!
Hi all. A couple of short notes for those using webglaze:
1. is the standard url to go to after you have created your account.
2. some have seen occasional server 500 errors. 1 helpful user helped identify potential cause: japanese unicode in filename.
Hi all. Quick couple of updates on (Web)glaze.
1. We put up a new document explaining how intensity values work in Glaze and how to use errors to guide your intensity parameters in Glaze. Page is here:
2. We increased max Glazes/day for all users.
Super quick note.
We'll be making a windows GPU version for glaze very soon. Initial tests show it's a ridiculous speedup. So if you're grinding through a backlog of images to glaze, might be worth waiting a day or so. It's a bigger download but a LOT faster.
WebGlaze update
Based on user feedback, we have tuned our default render time higher to better find protective Glazes under lower intensity constraints. New render is now equivalent to max render on the Glaze app, but adaptive (shortcuts when successful Glaze is found).
@Icemanmodeler
that "trash" is what Glaze makes AI models see when they try to train on the glazed image. It is designed to look like a version of the image that is stylistically completely distorted from the original.