Lecture 11 - DIAMOND | Diffusion as a World Model

In this lecture, we look at DIAMOND (Paper: https://arxiv.org/abs/2405.12399): instead of IRIS's discrete tokens, the world model denoises the next Atari frame directly in pixel space. We cover how the past frames are stacked along channels, how actions enter through adaptive group normalization, the EDM denoiser, and how an actor-critic agent is trained entirely inside the diffusion dream (1.46 human-normalized score on Atari 100k). Project: Build a DIAMOND diffusion world model on Atari and train an agent inside its dreams, then race your own world model on the World Model Arena leaderboa