2026-10-17 — views
GeoState: Learning View-Decoupled Geometry States for 3D World Modeling
Read this because This approach solves the entanglement of scene dynamics and camera variation, enabling more accurate and consistent 3D world predictions across different viewpoints.
GeoState introduces a view-decoupled geometry representation that separates scene dynamics from camera motion, improving predictive 3D world modeling.
The Challenge of View Dependence in 3D World Models
Recent advancements in 3D world modeling have increasingly relied on features extracted from geometry foundation models (GFMs). These features provide rich geometric priors that are essential for understanding and predicting 3D environments. However, a significant limitation persists: the view dependence of these features. This dependence entangles the actual dynamics of a scene with variations caused by the camera’s position and orientation. Consequently, predicting the evolution of the resulting state becomes considerably more difficult, as the model must disentangle true scene changes from mere observational shifts.
Introducing GeoState
To address this challenge, researchers have proposed GeoState, a novel view-decoupled geometry representation designed for predictive 3D world modeling. The core innovation of GeoState lies in its ability to separate scene-state evolution from camera-dependent observations. This separation allows the model to focus on the intrinsic dynamics of the environment rather than being confused by camera movement.
The architecture operates through a specific pipeline. First, a scene tokenizer aggregates multi-view GFM features into a unified, view-decoupled scene state. This state representation abstracts away the specific viewpoint from which the data was captured. Next, a geometry decoder queries this state using camera rays. This mechanism allows the system to reconstruct or predict what the scene should look like from any given perspective without re-encoding the view dependence.
Autoregressive Prediction and Scale Consistency
An autoregressive world model is then employed to predict future scene states. Crucially, camera motion is modeled separately from the scene state evolution. This decoupling offers significant practical advantages. The predicted states can be decoded under either predicted cameras or externally specified cameras, providing flexibility in simulation and rendering scenarios.
Furthermore, GeoState retains a single context-derived scale throughout the entire rollout. This means that the model does not require per-frame scale alignment, which is often a source of error and complexity in other methods. By maintaining a consistent scale derived from the initial context, the model ensures geometric consistency across time.
Performance and Generalization
Experiments demonstrate that GeoState outperforms prior 3D world models on the Waymo dataset. The model supports geometrically consistent cross-view querying, meaning that objects and structures remain consistent regardless of the viewing angle. Additionally, GeoState generalizes zero-shot to other datasets, including KITTI and nuScenes, indicating its robustness and adaptability to different real-world driving scenarios without additional training on those specific datasets.
Practitioner note
Developers should consider integrating GeoState’s tokenizer and decoder architecture when building predictive 3D environments to avoid scale alignment errors. Focus on decoupling camera motion from scene state prediction to improve long-term rollout consistency and cross-view accuracy.