Future Dynamic 3D Reconstruction: A 3D World Model with Disentangled Ego-Motion

2026-06-16

AI Summary

Existing generative world models struggle to disentangle ego-motion from scene dynamics, often resulting in geometrically inconsistent long-term predictions—such as object deformation or disappearance. This work proposes a novel 3D implicit world model that explicitly decouples agent ego-motion from environmental dynamics for the first time, leveraging inferred ego-motion as an implicit proxy for actions to enable geometrically consistent future 3D reconstructions. By integrating a teacher–student distillation strategy grounded in spatial commonsense priors from foundation models, the method achieves high-quality, zero-shot future prediction up to two seconds using only monocular video input. It significantly outperforms current approaches in both geometric consistency and long-term stability across multiple datasets.

Abstract

Forecasting the evolution of dynamic environments is crucial for autonomous agents. While generative world models have recently achieved high photorealism in 2D video synthesis by mixing ego-motion and environmental dynamics within the image plane, they exhibit physical inconsistencies, such as morphing or vanishing objects, especially over long time horizons. In this paper, we propose FR3D, a world model that predicts a persistent 3D latent representation for future dynamic 3D reconstruction. Unlike prior works that treat the world as a sequence of image-based features, FR3D explicitly decouples the 3D evolution of the scene from the agent's trajectory, treating the inferred ego-motion as a latent proxy for action. This disentanglement resolves the ambiguities between self-motion and world-motion, ensuring geometric consistency into the future. Furthermore, we introduce a teacher-student distillation strategy that leverages the spatial "common sense" of off-the-shelf foundation models, leading to robust zero-shot generalization. Extensive experiments demonstrate FR3D's strong performance for future dynamic 3D reconstruction from monocular observations across multiple datasets, even 2 seconds into the future. Project page: https://fr3d-wm.github.io.

Problem

• dynamic 3D reconstruction
• ego-motion disentanglement
• world model
• geometric consistency
• future prediction

Innovation

• disentangled ego-motion
• dynamic 3D reconstruction
• world model
• teacher-student distillation
• geometric consistency