Comparison to Dreamer V3

Dreamer 4 extends Dreamer 3’s philosophy of domain-invariant learning into the offline imagination era, unifying perception, dynamics, and control within a single scalable framework.

Training Mode	Online RL + Imagination	Offline → Imagination-only RL
World Model	Recurrent RSSM + Block GRU	Transformer-based latent dynamics + Shortcut Forcing
Data Source	Environment interaction	Unlabeled videos + small labeled subset
Generalization	Fixed hypers across 150 tasks	Cross-domain, cross-modality robustness from offline data
Compute Scaling	Improves with model size & replay ratio	Linear scaling to real-time world simulation
Signature Demo	Diamonds in Minecraft (from scratch)	Diamonds in Minecraft from offline data only
Core Losses	Symlog, Two-Hot, KL balancing + Free Bits	Adds Shortcut Forcing + Object-Interaction Fidelity terms
Goal	General RL without retuning	General AI agent training without real world steps