GeoState: View-Decoupled Geometry States for 3D World Modeling

Forum topic · 小凯 · 2026-10-17

Summary

GeoState is a view-decoupled geometry representation for predictive 3D world modeling, proposed to fix a key limitation of using geometry foundation model (GFM) features as world states: their view dependence entangles scene dynamics with camera-induced variation. The method introduces a scene tokenizer that aggregates multi-view GFM features into a view-decoupled scene state, and a geometry decoder that queries this state using camera rays. An autoregressive world model predicts future scene states, while camera motion is modeled separately, so predicted states can be decoded under predicted or externally specified cameras while maintaining a single context-derived scale across rollouts. Experiments show GeoState outperforms prior 3D world models on Waymo, supports geometrically consistent cross-view querying, and generalizes zero-shot to KITTI and nuScenes. Authored by Hongchang Liu, Jian Liu, Yujuan Ding, Liang Peng, Xun Jiang, Fei Ma, Xing Xu, and Yi Bin; paper available as arXiv:2610.22343.

论文概要

研究领域: CV 作者: Hongchang Liu, Jian Liu, Yujuan Ding, Liang Peng, Xun Jiang, Fei Ma, Xing Xu, Yi Bin 发布时间: 2026-10-15 arXiv: 2610.22343

Introduction

Recent 3D world models increasingly use features from geometry foundation models (GFMs) as predictive world states, benefiting from their rich geometric priors. However, the view dependence of these features entangles scene dynamics with camera-induced variation, making the resulting state evolution harder to predict.

Method: GeoState

GeoState is a view-decoupled geometry representation for predictive 3D world modeling that separates scene-state evolution from camera-dependent observation:

A scene tokenizer aggregates multi-view GFM features into a view-decoupled scene state.
A geometry decoder queries this state using camera rays.
An autoregressive world model predicts future scene states, while camera motion is modeled separately.

As a result, predicted states can be decoded under predicted or externally specified cameras, while maintaining a single context-derived scale across the entire rollout.

Results

Outperforms prior 3D world models on Waymo.
Supports geometrically consistent cross-view querying.
Zero-shot generalization to KITTI and nuScenes.