Glaze: Protecting Artists from Style Mimicry by Text-to-Image Models

• DOI: 10.48550/arxiv.2302.04222
• Publication Date: 2023-01-01
• Authors: Shawn Shan; Jenna Cryan; Emily Wenger; Hai-Tao Zheng; Rana Hanocka; Ben Y. Zhao
• Keywords: Robustness; Gloss (optics)

Recent text-to-image diffusion models such as MidJourney and Stable Diffusion threaten to displace many in the professional artist community. In particular, can learn mimic artistic style of specific artists after "fine-tuning" on samples their art. this paper, we describe design, implementation evaluation Glaze, a tool that enables apply "style cloaks" art before sharing online. These cloaks barely perceptible perturbations images, when used training data, mislead generative try artist. coordination with community, deploy user studies more than 1000 artists, assessing views AI art, well efficacy our tool, its usability tolerability perturbations, robustness across different scenarios against adaptive countermeasures. Both surveyed empirical CLIP-based scores show even at low perturbation levels (p=0.05), Glaze is highly successful disrupting mimicry under normal conditions (>92%) countermeasures (>85%).