Google Research Explores Creativity in Diffusion Models

Google Research has published a study on the creativity of diffusion models, revealing that their ability to generate novel data is due to the mathematical process of score smoothing. This process, a result of neural network training, allows models to interpolate between training data points, rather than simply memorizing them. The research highlights how this smoothing effect enables diffusion models to create new and plausible data samples, offering a clearer understanding of their generative capabilities. This finding demystifies the creative process of diffusion models, showing it as a predictable mathematical outcome.