AI image creators

 AI image creators such as Stable Diffusion and Google's Imagen were found to generate near-identical copies of images from their training data, including copyrighted images and photos of people, according to a pre-print paper.

AI image creators such as Stable Diffusion and Google's Imagen were found to generate near-identical copies of images from their training data, including copyrighted images and photos of people, according to a pre-print paper.   The Google-led research highlights how AI image generators — specifically diffusion models — can replicate original imagery that was created or posted by other people and used without their knowledge or permission. The researchers were able to show that Stable Diffusion and Imagen have the capacity to spit out imagery that's nearly identical to the original, with some distortion. This occurred even when the original photo appeared only once in the training data. However, generating this exact or near-exact imagery was rare. Out of 300,000 text prompts, Imagen replicated the exact image only 0.03% of the time. Still, the research does highlight how AI image generators may infringe on copyright. AI models are "supposed to generate novel images rather than spitting out a memorized version," Vikash Sehwag, one of the researchers, told Gizmodo. The study comes from researchers at Google, DeepMind, Berkeley, and Princeton.