Diffusion models have become a cornerstone of modern generative AI, producing images, audio, and other media that often appear strikingly original. Yet the inner workings of this 'creativity' have remained opaque. A new blog post from Google Research's Algorithms & Theory team takes a step toward demystifying the phenomenon by examining the mathematical and computational principles at play.
The post explores how diffusion models learn to reverse a noising process, gradually refining random noise into coherent samples. The analysis suggests that creativity may emerge from the model's ability to interpolate and combine features from its training distribution, rather than simply retrieving stored examples. This perspective helps distinguish between generation that is truly novel and output that is essentially a variation on existing data.
By clarifying these mechanisms, the research offers a framework for understanding both the strengths and limitations of diffusion models. It also raises questions about how we define creativity in machines, and whether the term can be meaningfully applied to algorithmic processes. The work is part of a broader effort within Google Research to build a more rigorous theoretical foundation for generative AI.