Generating videos that remain coherent when the camera moves dramatically is an open problem. According to a new arXiv paper, existing methods often fail to keep individual objects stable when viewpoints shift, leading to inconsistent representations. The paper introduces Totem, an approach built around object-centric 3D memory, to address this limitation.

The abstract describes Totem as a way to ground videos in a unified underlying world, rather than treating each frame independently. While the full details are not available in the abstract, the core idea is to use 3D memory centered on objects to preserve their identity and geometry across frames. This is a single-source summary, so no conflicting viewpoints are noted; the paper's claims are presented as the authors' own.