News
Adrian Baule wins EPSRC funding to investigate theoretical foundations of AI media generators
Centre for Complex Systems2 December 2025
Illustration of GDM transformations: An ordinary diffusion process maps structured data to random noise (left to right). The GDM reverts this process to create an image from noise (right to left). Source: NVIDIA blog (see below).
The UKRI Engineering and Physical Sciences Council (EPSRC) has awarded an £84,036 research grant for exploring Metastability in generative diffusion models to Adrian Baule, Reader in Applied Mathematics in the Centre for Complex Systems. The project will target the foundations of AI image, audio, and video generators such as DALL-E, Stable Diffusion, and Sora.
The underlying technology of such AI media generators is a generative diffusion model (GDM), which starts from random content and iteratively transforms it into structured data. Despite impressive practical results, the mechanism behind the success of GDMs is still poorly understood. Drawing on analogies with physical diffusion processes, the project aims to improve our understanding using tools from statistical physics. In particular, it will focus on metastability, a key ingredient for many physical, chemical, and biological phenomena. The insights are expected to enhance the performance of GDMs and help building more explainable, controllable, and responsible generative AI tools.
Image source: NVIDIA blog "Improving Diffusion Models as an Alternative To GANs, Part 2"
Email: l.dabelow@qmul.ac.uk
Updated by: Lennart Dabelow