A Chinese open-weight model has taken first place in the static section of WorldScore, a Stanford benchmark that tests how well AI systems generate consistent scenes as a virtual camera moves through them.
UniWorld-View scored 85.53, ahead of WorldScape-0.2(MoE) on 85.13. It did not lead the benchmark’s dynamic section, where the same rival model placed first on 76.23 against UniWorld-View’s 76.09 — a margin of 0.14 points, narrower still than the 0.40 that separates them on the static ranking. Scores are as of 27 July; the leaderboard is updated on a rolling basis.
The project does not describe itself as a world model. Its own repository calls it a large-baseline view synthesis system, though Chinese coverage of the result has consistently used the broader term. It was released by PKU-YuanGroup, a lab at Peking University led by Li Yuan, and generates new views of a scene from a single image or short video, letting users direct the virtual camera while the model attempts to preserve objects, geometry and visual detail across viewpoints.
The release is Apache-2.0 licensed and built on Alibaba’s open-weight Wan2.1-VACE video model, alongside components from Meta, Salesforce and an Nvidia research lab — a stack that is itself largely open-source and partly American. The repository recommends at least 60GB of GPU memory for inference, putting it beyond consumer hardware. No accompanying paper has been published.
The ranking comes with an important caveat. WorldScore accepts submissions two ways: teams can run the official evaluation code themselves and email the results to the maintainers, or send generated video samples for the WorldScore team to score. The leaderboard does not indicate which route a given entry took. Either way, the ranking reflects a submission process rather than an independent evaluation of the model by the Stanford group.
WorldScore evaluates camera control, visual quality and movement in generated scenes. It does not test whether a model can support an embodied agent interacting with its environment or accurately simulate real-world physics.
The result therefore shows strong performance in controllable scene generation, particularly for static environments, rather than establishing UniWorld-View as the best world model across all tasks.
