Glaze

The Glaze Project offers a free downloadable app called Glaze that uses machine learning to apply invisible pixel-level changes to artworks, causing AI image-generation models to misread an artist's style and fail to mimic it accurately. Artists apply Glaze to their images before posting them online, as a defensive measure against unauthorized style copying by systems like Midjourney and Stable Diffusion.

Glaze works by using style-transfer algorithms to identify the specific pixel features that define an artist's style, then computing minimal perturbations to those features that are imperceptible to the human eye but appear as a dramatically different style to AI models. The tool runs locally on the artist's computer (Windows and Mac), processes images one at a time, and produces a 'cloaked' image file the artist can post publicly. It was first released in March 2023 and has been downloaded more than 8.5 million times globally as of the project's most recent disclosure.

The Glaze Project deployed WebGlaze in August 2023, a free invite-only web service that lets artists apply the same style-cloaking protection as the desktop Glaze app directly from a phone, tablet, or browser, without needing a powerful personal computer. The Glaze team runs the processing on GPU servers at the University of Chicago.

WebGlaze accepts image uploads via a web form along with a strength parameter and email address; the image is routed to University of Chicago servers, processed by the Glaze algorithm on a GPU, and the cloaked result is emailed back to the user. Both the original and processed images are deleted immediately after delivery. Access is invite-only and restricted to human artists who do not use generative AI tools. WebGlaze was initially hosted on Amazon AWS but as of the most recent disclosure is run on a University of Chicago GPU server.

The Glaze Project offers a free downloadable tool called Nightshade that allows artists to apply invisible pixel-level changes to their images before posting online, causing AI image-generation models that train on those images to learn incorrect associations between objects and concepts. It was publicly released in January 2024 and has been downloaded more than 2.5 million times.

Where Glaze is defensive (preventing style mimicry of a specific artist), Nightshade is offensive: it embeds hidden pixel distortions that cause AI models to associate the wrong concepts with images — for example, training data showing dogs may register as cats to the model. The Glaze Project has described this as a 'data poisoning' approach intended to raise the cost to AI companies of scraping artists' work without consent. Nightshade was developed at the University of Chicago SAND Lab and peer-reviewed at the IEEE Symposium on Security and Privacy in May 2024. The team has stated they plan to integrate Nightshade as an optional feature within Glaze.