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

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.

Autonomous systems, such as robots or self-driving cars, need to predict how their surroundings evolve over time. Many current models forecast future video frames, but they mix two different sources of motion: the camera’s own movement and the movement of objects in the scene. This can lead to physically inconsistent predictions. We introduce FR3D, a world model that predicts the future directly in 3D rather than only in image space. It separates camera motion from scene motion, making future predictions more geometrically consistent. To train it, we transfer knowledge from a strong pre-trained 3D reconstruction model. Our experiments show that FR3D can forecast the evolution of a dynamic scene in 3D, even two seconds into the future, and for new datasets, given just a sequence of images.