UniWorld-View Tops Fei-Fei Li Team's World Model Leaderboard, Fully Open-Sourced with Ascend NPU Support

The world model leaderboard from Fei-Fei Li's team has a new top spot. UniWorld-View, jointly developed by Toozhan AI (兔展智能), Peking University, and Peng Cheng Laboratory, has claimed the top position. The model is already compatible with domestic Ascend NPUs and has been fully open-sourced with all code and weights available on GitHub.

UniWorld-View specializes in Novel View Synthesis (NVS), enabling AI to hallucinate unseen camera perspectives from a single image or monocular video. Users can capture a casual video or even just a single photo, and the model generates new-angle videos with freely controllable camera trajectories — push, pull, pan, tilt, or even a full 360-degree orbit around the scene, with precise camera pose control.

Traditional approaches such as NeRF and 3D Gaussian Splatting rely on the reconstruction paradigm, requiring multi-view inputs. When faced with monocular videos that have minimal coverage, these methods suffer from hollow artifacts or outright failure. Most generative approaches, meanwhile, simply feed camera poses as abstract condition vectors into diffusion models, lacking explicit 3D geometry — causing viewpoint drift and loss of control at larger baselines.

UniWorld-View's core technical innovation is its occlusion-aware point cloud rendering. The team identified a flaw in prior methods: conventional point clouds are collections of isolated points lacking connectivity or surface orientation cues. During large-baseline view changes, foreground textures get stretched across backgrounds like gum, or background pixels bleed through foreground objects. UniWorld-View uses Double-Reprojection to identify occluded regions and remove erroneous signals, and estimates normal vectors for each point — culling back-facing points from rendering entirely — providing clean geometric conditions for the diffusion model.

During generation, UniWorld-View employs a dual-stream conditioning strategy: one stream receives the point cloud render for positional accuracy, the other receives the source video for complete content. This ensures output aligns with the target viewpoint while preserving the original scene appearance. A further milestone is its end-to-end pipeline from monocular video to multi-view video to full 4D scene — by generating multiple novel views and feeding them into 4D Gaussian Splatting optimization, the model produces freely navigable 4D scenes from casual monocular capture.

Toozhan AI was founded by Peking University alumni and young leading researchers in computer vision from PKU. The company has developed the Open-Sora Plan and UniWorld series of visual AI models. Open-Sora Plan was the industry's first open-source text-to-video model and led global code citations in 2024, while the UniWorld series pioneered unified understanding-and-generation architectures. Toozhan previously released UniWorld-V2 and UniWorld-V2.5, matching GPT-Image-2's generation quality on infographics and text-dense scenarios.

The training data for UniWorld-View comes from YuanKong Intelligence, a Peking University-incubated startup. Importantly, the model's compatibility with domestic Ascend NPUs signals continued progress in China's domestic AI hardware infrastructure for large-scale visual model training and inference. Toozhan AI also disclosed plans to launch RabbitVis, a design-production tool integrating model capabilities, aiming to bring visual AI into commercial design workflows.

As competition in the world model space intensifies, UniWorld-View's combination of a top leaderboard ranking and fully open-sourced stack positions it to provide foundational capabilities for embodied intelligence, autonomous driving, and cinematic production — applications that demand rich 3D scene understanding.