Move the camera and the geometry features move too: splitting viewpoint out of the world state buys a 1.4-second forecast
Synopsis
Driving 3D world models often treat geometry-foundation-model features as future states, yet those features shift with camera viewpoint, so the model must predict scene evolution and camera motion at once. GeoState compresses multi-view features into a view-decoupled scene state that a camera ray queries only at decoding, while camera motion is modeled separately. On Waymo it beats compared world models at every reported horizon and transfers zero-shot to KITTI and nuScenes.
Interpretation
A 3D world model's predictive state no longer has to change with camera viewpoint, and one scene state can be decoded into depth under different cameras. Previously, geometry-foundation-model features served directly as predictive states, so camera motion alone shifted the features and the temporal model had to fit scene evolution and viewpoint change together. On Waymo with predicted cameras, the method beats DINO-Foresight, VGGT-World and FR3D on both depth metrics at every reported horizon, and the margin grows with horizon.
The scene state is aggregated from multi-view geometry features, and the query viewpoint enters only at decoding. A scene tokenizer aggregates frozen VGGT- multi-view features with 3D points, camera poses and ray encodings into a fixed-length state, and a geometry decoder retrieves that viewpoint's depth from Plücker rays of the query camera. The world model trains on roughly 140k Waymo windows, each with four context states and seven future states at 0.2 s intervals, and is evaluated on 100 held-out windows.
Camera motion is modeled separately, so a predicted state can be decoded under a predicted or an externally specified camera. The temporal model predicts only a residual update to the scene state, a separate camera predictor supplies the trajectory, and one context-derived scale is kept throughout the rollout. On 113 eligible KITTI drives and all 150 nuScenes validation scenes, without target-dataset adaptation, the method beats compared methods on both metrics at all horizons.
Perspective
The demonstrated scope is large-scale dynamic driving scenes: Waymo for in-domain evaluation and KITTI and nuScenes for zero-shot transfer, scored by AbsRel and δ<1.25 against LiDAR depth. The method addresses readers building 3D world models whose predictive states come from geometry-foundation-model features, with scale calibrated once on the observed context and kept through the rollout. The authors propose extending the framework to embodied settings where predictive scene states could support perception, planning and interaction, but give no results for that extension.
The authors note that VGGT- features depend on the other images in the same forward pass, and that relative differences in deep features are much larger than those of calibrated depth maps, so observation-derived target states are not purely geometric. Ablations show that extending state matching into autoregressive rollout degrades long-horizon performance, which the authors attribute to such context-dependent variation. In the cross-query consistency comparison, the observation-encoded reference uses target-time LiDAR calibration, making it a privileged reference.