Beyond Pixel Histories: World Models with Persistent 3D State

TL;DR

The paper introduces PERSIST, a world model that simulates a latent 3D scene to generate consistent, long-term video with spatial memory, enabling more realistic and controllable 3D environment synthesis.

Contribution

It presents a novel 3D-aware world model that maintains persistent spatial memory, improving 3D consistency and enabling environment editing from a single image.

Findings

Significant improvements in spatial memory and 3D consistency.

Enhanced long-horizon stability in generated videos.

Ability to synthesize diverse 3D environments from a single image.

Abstract

Interactive world models continually generate video by responding to a user's actions, enabling open-ended generation capabilities. However, existing models typically lack a 3D representation of the environment, meaning 3D consistency must be implicitly learned from data, and spatial memory is restricted to limited temporal context windows. This results in an unrealistic user experience and presents significant obstacles to down-stream tasks such as training agents. To address this, we present PERSIST, a new paradigm of world model which simulates the evolution of a latent 3D scene: environment, camera, and renderer. This allows us to synthesize new frames with persistent spatial memory and consistent geometry. Both quantitative metrics and a qualitative user study show substantial improvements in spatial memory, 3D consistency, and long-horizon stability over existing methods, enabling coherent, evolving 3D worlds. We further demonstrate novel capabilities, including synthesising diverse 3D environments from a single image, as well as enabling fine-grained, geometry-aware control over generated experiences by supporting environment editing and specification directly in 3D space. Project page: https://francelico.github.io/persist.github.io