Frequently Asked Questions

First, here are a few basic things to note about Nightshade. Please also see the FAQ page for Glaze, since many of the basic questions are answered there first.
  • Image-specific: Nightshade works to change your input image, so every image will produce a different result. In addition, there is a randomized component in Nightshade (just like Glaze), so that each run with the same input image will not produce the same output. If you shade an image and don't like it, do it again and you will likely see something fairly different.
  • Efficacy against different AI models: Like Glaze, Nightshade operates with open source AI models as a guide in its computation. That means it is most effective on Stable Diffusion models. Transferability generally means there will be fairly strong effects on other diffusion models, but the precise target might not look the same, and the strength per image might be weakened (i.e. it might require more shaded images to have the same net effect.
  • Robust against removal: Nightshade poison is computed in a similar way as Glaze. Like Glaze, it cannot be easily removed from the artwork using normal (non-ML) transformations/edits (e.g., sharpening, blurring, denoising, downsampling, stripping of metadata, etc.).

Does Nightshade scrape, copy, train on your art?
Absolutely not. Nightshade is a tool to empower artists. Like Glaze, Nightshade is designed to work perfectly fine in offline mode. So you can install it, turn off your wifi and unplug any networks, then run Nightshade, and quit the program before you turn the wifi back on. Since late 2022, our team has been working nonstop to fight for the rights of human artists. From day 1, we have made it clear that our goals do not involve money or profit. There is no hidden startup (plenty of VCs have approached us, we are not interested); there is no selling IP to some larger tech company; there is no model training. If we need funds to expand engineering or to help deploy Glaze/Nightshade at larger scale, we will seek donations from foundations for the arts as a non-profit.

What about artists who use mobile devices?
Yes, we are aware that many, perhaps most artists do not have easy access to powerful GPUs on desktop machines. But we do not have the time and resources to build mobile apps. Instead, we will work to integrate Nightshade into the WebGlaze webservice. That will run on any mobile device with a web browser, and is much easier than a mobile app.

I gave a Nightshaded/Glazed image to GPT/BLIP and it recognized it perfectly. Does this mean Nightshade/Glaze failed? No, it does not. Nightshade and Glaze both target image generators, which are built on diffusion architectures. Image classification, which is what you get when you ask a model to tell you what is in an image, is a completely different task. The normal properties of transferability that allow attacks or perturbations targeting one model to affect another similar model, generally does not extend to models that perform different tasks. Today's prompt extraction tools are not traditional DNN classifiers, but are still different enough architecturally from image generators to completely break transferability. To put it plainly, Nightshade and Glaze are designed to NOT affect those models. If a model identifies the contents of your shaded/glazed image perfectly, then that is correct behavior. In fact, consider the opposite. If anyone can identify a shaded image not as the original, but as the shade target, then this is a super fast and easy way to identify/filter out all Nightshaded images, and Nightshade would not be useful.

Is Nightshade or use of Nightshade illegal? Can users be found liable or sued by AI companies? I am not a lawyer, so I gone 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. One accurate analogy for using Nightshade is like putting hot sauce in your lunch that is clearly labeled "DO NOT EAT/SCRAPE." Someone stealing your food is doing so against your express wishes, and if they get sick from the hot sauce, you are not liable. 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. Contrary to some rumours/misinformation online, Nightshade is not a virus, it does not infect/affect any model against their wishes. And Nightshade can only have an effect on diffusion models, and only when a lot of shaded images are trained. Nightshade will not affect large language models, or any non-generative AI systems (medical imaging, facial recognition, self-driving cars etc...)

What kind of image quality can we expect from Nightshaded images? 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. We have not yet added lots of Nightshaded image samples online. We will get around to that as we have some time to ask for permission from our many artist friends who have deployed Nightshade.

What about Mac GPU, Mac Intel, non-NVidia GPU support? What about Linux machines? Other versions?
Yes we would love to do all this to support different platforms. Unfortunately now there are quite a few Glaze and Nightshaded related projects running in parallel in our lab, and our time is spread very thin. We did plan on an open source version of some sort. That will probably come, but it is lower priority compared to expanding GPU support for more artists.

What about that pesky NVidia GTX 1660/1650/1550 bug? Yes, if you are getting a whole black image after applying Nightshade, it's probably because you have one of the NVidia GTX1660/1650/1550 GPUs. There is a bug with PyTorch ML library on this GPU that makes it fail half-precision floating point instructions. It is really hard for us to fix bugs that are in PyTorch. We will eventually fix this, most likely by reimplementing Nightshade/Glaze on Tensorflow. But that is an engineering-heavy process, and we do not have the engineering power to deal with it in the short term.

Isn't it true that Nightshade can be broken/bypassed by pixel-smoothers?
Actually, No. Ever since Glaze was released in March 2023, there have been misinformation campaigns about how smoothing out the visible artifacts on a Glazed image will break Glaze. The pixel cleaner posted on Github (AdverseCleaner) does not work, and the author of the tool admitted it in an update post on that Github page. Despite this, the misinformation campaigns continued. Finally, the author of the adverseCleaner tool simply took down his page altogether off of Github.


Why does pixel smoothing not break Glaze/Nightshade? It might be helpful to understand that Glaze and Nightshade do not just tweak a few pixels, and those magical pixels are the visible artifacts you see. No, as it turns out, Glaze/Nightshade actually change the large majority of all pixels in the image (80%+). So the entire image is being altered, and the visible artifacts are just the tip of the iceberg in terms of changes. Smoothing them out only changes a very small portion of the image changes. One analogy I sometimes use is that using pixel smoothing on a Glaze/Shaded image is like your house has been hit by an earthquake and moved three blocks away from its original location, but you go into the dining room and rearrange the chairs back into their rightful place.

How can I help support Nightshade and Glaze?
Thank you so much for thinking of us. It is important to us that we not only continue to provide Glaze to visual creators for free, but also extend its protective capabilities. If you or your organization may be interested in pitching in to support and advance our work, you can donate directly to Glaze via the Physical Sciences Division webpage, click on "Make a gift to PSD" and choose "GLAZE" as your area of support (managed by the University of Chicago Physical Sciences Division).

Background Image: Mermaid, Emily Odette, Jingna Zhang