Video diffusion models have long struggled to maintain visual continuity when the virtual camera moves or revisits previous locations, often leading to visual glitches where the environment changes unexpectedly. Traditional methods usually choose between explicit 3D structures, which are rigid and struggle with moving objects, or implicit memory, which often fails to follow precise camera paths accurately.

Researchers have introduced MosaicMem (Mosaic Memory), a hybrid spatial memory system that bridges these two approaches. By "lifting" small sections of an image (patches) into a 3D coordinate system, the model can accurately place and retrieve visual information based on where the camera is looking. This patch-and-compose technique allows the AI to preserve stable background structures while naturally filling in (inpainting) new or moving elements, ensuring the world remains cohesive over time.

The system employs sophisticated alignment techniques to fuse 3D geometry with the AI's internal generation process without requiring expensive retraining. This breakthrough enables advanced capabilities such as minute-long navigation through virtual spaces and complex scene editing. Users can now expect video world models that act more like consistent simulators rather than a series of disconnected clips, paving the way for more immersive AI-generated environments.

Video diffusion models have long struggled to maintain visual continuity when the virtual camera moves or revisits previous locations, often leading to visual glitches where the environment changes unexpectedly. Traditional methods usually choose between explicit 3D structures, which are rigid and struggle with moving objects, or implicit memory, which often fails to follow precise camera paths accurately.

Researchers have introduced MosaicMem (Mosaic Memory), a hybrid spatial memory system that bridges these two approaches. By "lifting" small sections of an image (patches) into a 3D coordinate system, the model can accurately place and retrieve visual information based on where the camera is looking. This patch-and-compose technique allows the AI to preserve stable background structures while naturally filling in (inpainting) new or moving elements, ensuring the world remains cohesive over time.

The system employs sophisticated alignment techniques to fuse 3D geometry with the AI's internal generation process without requiring expensive retraining. This breakthrough enables advanced capabilities such as minute-long navigation through virtual spaces and complex scene editing. Users can now expect video world models that act more like consistent simulators rather than a series of disconnected clips, paving the way for more immersive AI-generated environments.