eloialonso/diamond

Builds diffusion-based world models capable of multi-step visual rollouts in complex environments like Atari and Counter-Strike, enabling both policy training entirely within imagination and interactive play with human control. Uses Hydra for configuration management and integrates with Hugging Face Hub for pretrained model distribution, supporting training resumption, W&B logging, and model-free RL baselines for comparison. Implements configurable diffusion sampling (denoising steps, stochasticity) with playable visualizations that allow switching between learned policies, human control, and real environment interaction across imagination horizons.

1,988 stars. No commits in the last 6 months.