Researchers have found that the perceived creativity of diffusion models stems from their deterministic architecture, with imperfections in the denoising process contributing to their ability to produce seemingly original images.

Recent research indicates that diffusion models' focus on small image patches through local, patch-focused processes leads to the generation of novel, seemingly creative combinations, providing potential insights into human creativity and brain function.

Experiments with a simplified system called the equivariant local score (ELS) machine show that local mechanisms can predict the outputs of trained diffusion models with about 90% accuracy, reinforcing the idea that local processes are central to their creative capabilities.

Diffusion models, which are used in popular image generation tools like DALL·E and Stable Diffusion, create images by transforming noise into coherent visuals, a process that unexpectedly results in creative and novel outputs.

A mathematical model developed by physicists Mason Kamb and Surya Ganguli suggests that the 'creativity' of these models arises from local processing features such as locality and equivariance, which focus on individual image patches rather than broader context.