Diffusion Models: From Foundations to a Working Application

Diffusion Models: From Foundations to a Working Application is a Tokyo AI technical session at KERNEL HONGO in Bunkyo, covering diffusion models from flow matching and core design choices through diffusion language models to applications in discrete domains and biology. It is aimed at ML practitioners, AI engineers, researchers and research students who want technical explanations grounded in concrete examples. Xuan Zhang (SmartNews) introduces diffusion and flow matching with an emphasis on physical intuition and live demos rather than derivations. Federico Berto (Radical Numerics) explains how diffusion language models differ from autoregressive LLMs, shares lessons from scaling the open-source RND1 model, and shows how genome foundation models like Omnii treat DNA and proteins as language. A third talk on diffusion LLMs has its speaker still to be announced. Talks are followed by an hour of networking. November 18 is the preferred date, with November 20 as a fallback depending on speaker availability.