worldmodels.fyi
← Catalog
PaperUniversity of Geneva / Edinburgh / Microsoft Research · 2024

Diffusion for World Modeling: Visual Details Matter in Atari (DIAMOND)

Trains RL agents inside a diffusion world model, showing that high-fidelity visual detail materially improves downstream policy performance on Atari.

Summary

DIAMOND (DIffusion As a Model Of eNvironment Dreams) replaces discrete-latent world models with a diffusion-based next-frame predictor, letting agents train in a visually faithful imagined environment. It sets a strong Atari 100k result and shows visual detail preserved by diffusion matters for control.

Metadata

Authors
Alonso et al.
Year
2024
arXiv
2405.12399
Venue
NeurIPS 2024 (Spotlight)

Relationships

Introduces

DIAMOND
#diffusion#rl#atari