Image generators caught photocopying their homework, researchers say
Ourtake
Stable Diffusion reproduced 103 training images. Nobody was shocked, except possibly the 103.
What happened
Researchers from Google, DeepMind, ETH Zurich, Princeton and UC Berkeley found that Stable Diffusion and Google's Imagen can memorize and reproduce individual training images. They extracted more than 100 near-identical replicas from Stable Diffusion, ranging from personally identifiable photos to copyrighted logos. The team recommends deduplicating training data and avoiding diffusion models where privacy is a major concern, such as medicine.