✦
Summary
DreamX-World is a general-purpose interactive world model for creating diverse, high-fidelity worlds that can be explored, controlled, and transformed through actions and event prompts. It moves beyond passive video generation by supporting controllable navigation, third-person generation, promptable events, and persistent long-horizon rollouts.
Key Features
Creates controllable worlds that can be explored through action inputs.
Supports long-horizon autoregressive video rollouts over hundreds of frames.
Uses geometry-guided world memory to retrieve non-local visual evidence.
Supports first-person navigation and coherent third-person world generation.
Handles realistic, imaginative, game-like, science-fiction, and stylized environments.
Responds to prompt-driven world events beyond camera movement alone.
Trained on Unreal Engine data, gameplay footage, and real-world videos.
Links to arXiv, public code, and Hugging Face model resources.
About DreamX-World
The model is trained from a scalable data engine that mixes Unreal Engine data, gameplay footage, and real-world videos, with camera estimation, filtering, and curated distributions used to teach realistic dynamics. Its training pipeline progressively learns world dynamics, fine-grained action control, event response, reinforcement-learning improvements, and distilled efficient inference.
DreamX-World is useful for embodied AI, game-like simulation, interactive video generation, and agents that need spatially coherent environments over many frames. World memory and geometry-guided retrieval help the model preserve layout and object identity when the camera revisits a region.