On Copyright Risks of Text-to-Image Diffusion Models

Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, Kenji Kawaguchi

Keywords: Copyright imfringement, diffusion models, text-to-image generation

Abstract: Diffusion models excel in generating high-quality images from text prompts but often replicate elements from their training data, raising copyright concerns. While recent studies focus on direct, copyrighted prompts, our research examines subtler infringements triggered by indirect prompts. We introduce a data generation pipeline to systematically study copyright issues in diffusion models, replicating visual features using seemingly irrelevant prompts for T2I generation. Testing various models, including Stable Diffusion XL, our results reveal a widespread tendency to produce copyright-infringing content, highlighting a significant challenge in this field.

Submission Number: 7