MoWorld is a real-time interactive world model for NPUs, positioned as a high-frame-rate real-time interactive world model, with the entire pipeline from training and inference to deployment built on a domestic platform, aiming to achieve high-frame-rate, low-cost, continuously controllable video world generation.
Purpose and Positioning The core purpose of MoWorld is to generate interactive video worlds and support continuous control. According to official materials, it can be used for indoor scene reconstruction, exploring generated worlds through continuous camera control, and rapid generation of video worlds for large scenes, realistic lighting, and director-grade shots, serving scenarios such as storyboarding, previsualization, and proof of concept.
Training Data and Quality Screening MoWorld's training data covers a variety of scenarios in both the real world (indoor, driving, etc.) and virtual worlds (games, anime, sci-fi films, etc.), balancing richness and generalization capability. At the same time, through multi-dimensional quality screening such as 3D geometric consistency, trajectory accuracy, and multi-view stability, as well as key processes including geometric completion, dynamic region filtering, and trajectory verification, high-quality training assets are selected from massive amounts of data, providing a data foundation for low-cost, high-performance training.
Real-Time Inference Optimization MoWorld decomposes the real-time inference problem into three levels of collaborative optimization: pipeline level, parallel level, and operator level. It achieves efficient VRAM utilization in resource-constrained scenarios, accommodating massive model parameters, communication buffers, and inference states; at the same time, it compresses end-to-end latency, reflects user control signals in generated frames in real time, and continuously outputs in a streaming manner.
Main Capabilities and Usage Official materials show that MoWorld supports continuous camera control to explore generated worlds, supporting low-latency roaming, camera control, and interactive content experiences; using its generated videos for indoor scene reconstruction achieves high target precision, stable reconstruction, and good spatial consistency. Regarding specific installation commands, code examples, API calling methods, platform entry points, and pricing models, the materials do not currently provide them.
Applicable Scenarios Including but not limited to: indoor scene reconstruction, interactive roaming and camera control, storyboarding and previsualization, proof of concept, and video world generation for large scenes, realistic lighting, and director-grade shots.
Limitations The materials do not provide MoWorld's specific system requirements, hardware compatibility list, model parameter scale, license agreement, usage threshold, or pricing information; the above content is not currently provided and should be subject to the official page.
