MosaicMem: Hybrid Spatial Memory for Controllable Video World Models
This addresses the problem of maintaining spatial consistency in video generation for applications like navigation and editing, representing an incremental improvement over existing methods.
The paper tackled the bottleneck of spatial memory in video diffusion models for consistent world simulation under camera motion and interventions, proposing MosaicMem, a hybrid spatial memory that improved pose adherence compared to implicit memory and enhanced dynamic modeling over explicit baselines.
Video diffusion models are moving beyond short, plausible clips toward world simulators that must remain consistent under camera motion, revisits, and intervention. Yet spatial memory remains a key bottleneck: explicit 3D structures can improve reprojection-based consistency but struggle to depict moving objects, while implicit memory often produces inaccurate camera motion even with correct poses. We propose Mosaic Memory (MosaicMem), a hybrid spatial memory that lifts patches into 3D for reliable localization and targeted retrieval, while exploiting the model's native conditioning to preserve prompt-following generation. MosaicMem composes spatially aligned patches in the queried view via a patch-and-compose interface, preserving what should persist while allowing the model to inpaint what should evolve. With PRoPE camera conditioning and two new memory alignment methods, experiments show improved pose adherence compared to implicit memory and stronger dynamic modeling than explicit baselines. MosaicMem further enables minute-level navigation, memory-based scene editing, and autoregressive rollout.