Data Collection via Agent Play: Since we cannot collect human gameplay at scale, as a first stage we train an automatic RL-agent to play the game, - persisting it's training episodes of actions and observations, which become the training data for our + persisting its training episodes of actions and observations, which become the training data for our generative model.
Training the Generative Diffusion Model: We re-purpose a small diffusion model, Acknowledgements