Video World Models with Long-term Spatial Memory

Tong Wu (Beijing Institute for General Artificial Intelligence) · Shuai Yang (Nanyang Technological University) · Ryan Po (Stanford University) · Yinghao Xu (Stanford University) · Ziwei Liu (Nanyang Technological University) · Dahua Lin (TTI Chicago) · Gordon Wetzstein (Stanford University)

3d memory mechanismsautoregressive models context lengthcustom datasetsforgetting mechanismsgeometry-grounded frameworkhuman memory mechanismsinformation retrieval long-term memoryquality evaluationscene consistencyspatial memorytemporal contextvideo frame generationworld models

Emerging world models autoregressively generate video frames in response to actions, such as camera movements and text prompts, among other control signals. Due to limited temporal context window sizes, these models often struggle to maintain scene consistency during revisits, leading to severe forgetting of previously generated environments. Inspired by the mechanisms of human memory, we introduce a novel framework to enhancing long-term consistency of video world models through a geometry-grounded long-term spatial memory. Our framework includes mechanisms to store and retrieve information from the long-term spatial memory and we curate custom datasets to train and evaluate world models with explicitly stored 3D memory mechanisms. Our evaluations show improved quality, consistency, and context length compared to relevant baselines, paving the way towards long-term consistent world generation.