Original title: Investigating Copyright Issues of Diffusion Models under Practical Scenarios
Authors: Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, Haonan Wang, Kenji Kawaguchi
The article delves into copyright concerns linked with generative models, particularly focusing on diffusion models. Past studies mainly centered on complete image replication, overlooking nuanced copyright infringement aspects. They used prompts closely resembling copyrighted topics. However, our work ventures beyond, examining partial copyright infringement, where specific image segments are treated as copyrighted. We develop a data generation pipeline to create diverse datasets for copyright exploration in diffusion models. These datasets highlight infringement patterns across various models, including the latest Stable Diffusion XL. Evaluations on generated data underscore the prevalence of copyright-infringing content, shedding light on the need for nuanced copyright understanding and mitigation strategies within generative models.
Original article: https://arxiv.org/abs/2311.12803