Generative diffusion model with inverse renormalization group flows

Diffusion models have recently emerged as one of the most powerful frameworks for generative modeling, achieving remarkable success in a wide range of domains, including image generation, audio synthesis, and scientific data generation. However, despite their empirical success, conventional diffusion models often require many denoising steps and do not explicitly exploit the multiscale structure naturally present in various types of data. This limitation motivates us to ask whether ideas from the renormalization group (RG), which is designed to describe scale-dependent effective degrees of freedom, can provide a useful principle for constructing more efficient generative models.

In this talk, I will present our recent work on renormalization-group diffusion models (RGDMs) [1], a generative framework that connects diffusion models with RG flows. By establishing a correspondence between diffusion dynamics and exact RG flow equations, we construct a diffusion model whose reverse process generates data in a coarse-to-fine manner, thereby effectively reversing an RG flow.

I will first introduce the theoretical formulation of RGDMs and explain how the RG perspective leads to a coarse-to-fine generative process. I will then present numerical results in protein structure prediction and image generation, where RGDMs improve sample quality and/or sampling efficiency compared with conventional diffusion models. Finally, I will discuss possible extensions and open questions, including broader applications of RG-inspired generative modeling.