Researchers Attempt to Uncover the Origins of Creativity in Diffusion Models

Stanford researchers developed a mathematical model called the "equivariant local score (ELS) machine" to explain diffusion models' creativity in image generation. The model incorporates translational equivariance and locality biases, suggesting diffusion models create new images by combining local patches from training images. With about 90% accuracy in predicting outputs, the ELS machine offers insights into how these AI systems generate creative and occasionally flawed images.