UniWorld-View Tops WorldScore: A Chinese Open-Source World Model for Controllable Novel Views

Lead

A new Chinese open-source project has drawn attention in the world model race. UniWorld-View, developed by Tuzhan Intelligence together with Peking University and Pengcheng Laboratory, has taken the top position on the WorldScore leaderboard. Its promise is easy to understand: provide one image, or a casual monocular video, specify a camera path, and the system generates a new-view video with controlled push, pull, pan, translation, or even a 360-degree scene orbit.

The broader point is that UniWorld-View uses one model and one codebase for both single-image 3D generation and video-based 4D generation. According to the source material, it has also been adapted to domestic Ascend computing hardware, with both code and weights released.

Key points

• Aiming at large-baseline novel view synthesis: Classic NeRF and 3DGS pipelines rely heavily on multi-view coverage. With casual monocular footage, insufficient viewpoints often lead to holes and artifacts. Pure generative methods can hallucinate missing views, but without explicit 3D constraints they may drift or fail under large camera changes.
• Occlusion-aware point-cloud rendering: UniWorld-View follows a geometry-plus-generation route, but focuses on a major weakness of point-cloud conditions: tearing between foreground and background, and leakage from the wrong side of an object.
• Double reprojection for occlusion masks: The method projects source pixels to the target view and then reprojects them back. Pixels that cannot return reliably are treated as occluded regions and marked with masks, leaving holes for the generative model instead of feeding it misleading textures.
• Normal filtering to reduce back-side leakage: By estimating point normals and comparing them with the viewing direction, the renderer filters points facing away from the target camera, reducing cases where front textures incorrectly appear from behind.
• Dual-stream conditioning: A geometry stream provides the target-view structure through rendered point clouds and masks. An appearance stream, built around the Ref-DiT module, uses the source video as reference material so the model can retrieve details and maintain visual consistency.
• A path toward 4D reconstruction: The pipeline first freezes time at the initial frame to create a static orbit of appearance anchors. It then generates synchronized multi-view videos from fixed viewpoints and feeds them into 4DGS optimization to obtain a dynamic scene that can be explored from free viewpoints.

Why it matters

UniWorld-View is not merely a “make a video from an image” system. Its technical value lies in combining geometric constraints with generative completion: unreliable regions are explicitly masked, wrong back-facing points are removed, and the diffusion model fills the missing content under stronger spatial guidance.

This could lower the barrier for turning ordinary monocular footage into 3D or 4D content. Potential applications include film previsualization, game assets, product display, spatial walkthroughs, and visual design workflows. The open release of code and weights should also help researchers reproduce results and compare approaches more directly.

Still, topping a benchmark does not mean every real-world scenario is solved. Large viewpoint changes, complex dynamic occlusion, and fine structural consistency remain important stress tests. UniWorld-View should be seen as a notable step toward controllable 4D scene generation from sparse visual input.

Source: QbitAI