Translations:Diffusion Models Are Real-Time Game Engines/73/en

We compare training on agent-generated data to training on data generated using a random policy. For the random policy, we sample actions following a uniform categorical distribution that doesn’t depend on the observations. We compare the random and agent datasets by training 2 models for 700k steps along with their decoder. The models are evaluated on a dataset of 2048 human-play trajectories from 5 levels. We compare the first frame of generation, conditioned on a history context of 64 ground-truth frames, as well as a frame after 3 seconds of auto-regressive generation.