Xiaomi EV has officially unveiled its Xiaomi EV World Model, a new framework for autonomous driving that deeply integrates 3D reconstruction and video generation for the first time. In autonomous driving simulation, traditional techniques often keep reconstruction and generation separate. Reconstruction modules can restore a scene but can't predict changes, while generation modules can forecast the future but suffer from drift and distortion over long sequences. The team proposes the JointWM architecture, which uses 3D geometry as a physical skeleton to anchor the scene, then employs a generation module to fill in visual details and predict unobserved areas. It sets new state-of-the-art performance across multiple benchmarks including Waymo and nuScenes.
Specifically, the reconstruction module, WorldRec, abandons the conventional per-pixel paradigm in favor of sparse 3D query points for scene representation, incrementally fusing them into a cross-view 4D Gaussian spatial skeleton, enabling a 10-second video to be reconstructed in 10 seconds. Based on the geometric prior provided by the reconstruction module, the generation module WorldGen is constrained by the skeleton's physical boundaries and only responsible for generating plausible lighting and textures. For content beyond those boundaries — such as future frames and blind spots — the generation module uses a two-stage temporal training and distribution-matching distillation mechanism to make physically grounded predictions. The entire architecture runs on an H20 GPU, achieving 0.19 seconds for a single view and 0.46 seconds for three views, and supports video generation up to one minute long.
On the Waymo reconstruction accuracy test, the solution achieves a PSNR of 28.48, and it maintains a leading position in zero-shot generalization on nuScenes. In terms of generation efficiency, it is 5.6x faster than the autoregressive baseline Epona, and its spatiotemporal coherence ranks among the top in its class. The research has already been deployed in three scenarios at Xiaomi EV: delivering over 100,000 high-quality synthetic data segments for perception model training, building high-fidelity closed-loop simulation environments to reproduce long-tail road conditions, and launching a driver assistance academy that uses generated videos to guide user operations.