DreamX World

DreamX-World is a general-purpose interactive world model from the DreamX team at Alibaba (Amap). It generates worlds that can be explored through camera/action inputs and altered with text event prompts, going beyond passive video generation. It covers realistic settings (indoor scenes, urban streets, landscapes, architecture) as well as fantasy, game-like, sci-fi and stylized environments, and supports both first-person and third-person views.

Training uses a mix of Unreal Engine data, gameplay footage and real-world videos, in stages: world dynamics and action control first, then open-ended event response, then reinforcement learning, followed by forcing and distillation for efficient inference. Long autoregressive rollouts over hundreds of frames use rollout correction to reduce style and color drift, and geometry-guided memory retrieval brings back earlier frames when the camera revisits a region, helping keep layout and object identity consistent.

According to a YouTube creator's summary, the released 5B version is based on Wan 2.2 and is about 21 GB, with a 14B version planned; the creator also reports it as more consistent than LingBot World and HY World Play. The project page says the paper, model and code are provided to the research community, but no repository, Hugging Face or paper links, or license, were in the supplied material. The authors list real-time generation of long-duration worlds as future work.