"Glaze 2: Launch of Enhanced Anti-AI Scraping Tool for Artists with Upcoming Video Defense Features"

In February 2023, a team of researchers at the University of Chicago, led by computer science professor Ben Zhao, introduced Glaze—a free software tool that employs machine learning to subtly modify the pixels of user-uploaded artwork. This alteration changes how AI art generators perceive an artwork’s style without impacting its appearance to human viewers. For instance, a hand-drawn illustration might be interpreted by an AI as a watercolor piece or CGI art.

The main objective of Glaze was to assist artists in combating AI models that mimic their artistic styles while ensuring their original work remains intact. According to the research paper by Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y. Zhao, the tool aims to safeguard artists from loss of commissions and damage to their reputation, emphasizing that their styles are integral to their identity. The unauthorized use of their unique styles is likened to identity theft.

Artists can create an image, apply Glaze, and then upload the altered version online, ensuring that AI models cannot easily train on their work. The glazed artwork appears similar to the original to human viewers, but to an AI, it is fundamentally different. While effective, Glaze is most beneficial for artworks not previously scraped by AI.

By March 2024, Glaze had amassed over 2.3 million downloads, along with attention for the open-source follow-up tool, Nightshade, which aims to “poison” AI models trained on artists' creations without consent.

A year later, the Glaze Project team announced Glaze 2, designed for faster use and enhanced protection against advanced AI models, including Stable Diffusion XL, an open-source text-to-image generator that can be fine-tuned to emulate specific artists.

Improvements in Glaze 2

The original Glaze tool took an average of 1.2 minutes on a Titan RTX GPU and 7.3 minutes on an Intel i7 CPU to process a single artwork. Despite its effectiveness—with over 92% success in preventing style mimicry—Glaze 2 boasts significant speed improvements, processing times now enhanced by 50% to 500% based on the user's hardware. For example, some older GPUs have improved from 4 minutes to 2 minutes per image, while Mac users with M1-M3 processors can experience increases up to 5X.

Glaze 2 also offers stronger protection against models like SDXL, especially for smooth-surface art styles such as anime and cartoons. Zhao noted in a recent email that Glaze 2 creates more disruptive artifacts, making it harder for AI to replicate the protected work.

When asked about protections against style mimicry from other AI generators like OpenAI’s DALL-E 3 or Midjourney, Zhao emphasized that Glaze’s primary focus is on Stable Diffusion, as it is the predominant model under user control for mimicking artists’ styles.

Future Developments

The University of Chicago’s Glaze Project is fully invested in the new version, with Zhao confirming that Glaze 1 will no longer be available for download. "Glaze 2 is a drop-in replacement that should exceed Glaze 1 in every aspect," he stated.

Additionally, the team is exploring extending Glaze-like protection to video formats, as announced on their X account. They are conducting user studies to assess how well AI can replicate video frames and how effectively Glaze’s new version can disrupt this mimicry.

If Glaze 2 performs as well as its predecessors, it is poised to be a valuable asset for artists seeking to protect their works from unauthorized AI scraping.