AMAP (Alibaba) DreamX-World
DreamX-World, from AMAP (Alibaba Group), is a general-purpose interactive world model that generates diverse, high-fidelity worlds you can explore, control and transform with event prompts. Trained on Unreal Engine data, gameplay footage and real-world video with camera estimation, it uses a progressive pipeline (fine-grained action control → open-ended event response → RL) and geometry-guided memory to keep scenes consistent when the camera revisits an area. Open-sourced (DreamX-World-5B / 5B-Cam) and served in real time on Reactor.
Overview
| Status | Research |
|---|---|
| Access | Open weights / code |
| Released | 2026 |
| Inputs | image, text prompt, event prompts, camera control |
| Outputs | generated interactive video / world |
| Best for | promptable world events, long-horizon exploration, first- & third-person generation, camera-controlled navigation, research & interactive media |
Why it matters
Promptable, compositional world events plus long-horizon memory push real-time world models toward genuinely interactive, editable environments — and Reactor's serving makes it reachable via API.
Roamscape use
Tracked in Roamscape's model hub. Open weights are on HuggingFace / ModelScope; not yet available via a hosted real-time API, so it isn't runnable in-app yet.
Strengths
- single + compositional promptable world events
- geometry-guided memory (scene persistence on revisit)
- first- and third-person generation
- precise camera control
- open weights + published tech report
Limitations
- autoregressive long-horizon inference is compute-heavy
- not yet exposed on a hosted real-time API (open weights only)
- requires self-hosting the open-source checkpoints to run today
Sources
Related models
General-purpose real-time interactive world model.
Real-time streaming video-to-video editing at 24 FPS — live on Roamscape via Reactor.
Next-gen real-time world model: minute-scale horizons, sub-second latency, camera + action control.