Geometry as Address: Routing Attention to Visual Memory for Long-Horizon Camera-Controlled Video Generation

Geometry as Address. Routing Attention to Visual Memory for Long-Horizon Camera-Controlled Video Generation. By Zesong Yang and twelve others, from Zhejiang University, the Hong Kong University of Science and Technology in Guangzhou, and one other institution. Published on the twenty-eighth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Generating long camera-controlled video means recovering content that was seen earlier, out of a visual history that keeps growing. Existing approaches take one of two routes. Some search the historical context implicitly, which makes memory access inefficient. Others reconstruct that history into a persistent 3D memory, which lets geometric errors accumulate over time.

GEAR rests on a division of labour. Geometry does not need to explain the scene. It only needs to determine where visual memory should be read from, while attention decides what should be recovered. Short for Geometry-Enabled Attention Routing, GEAR uses geometry as an explicit address for visual memory at the level of individual tokens. Geometry becomes the catalogue entry that says which shelf to look on, rather than a copy of the book.

Historical observations are never fused into a persistent global 3D representation. GEAR keeps them as frame latents and uses the geometry of each frame only to establish token-level correspondences with the target views, which avoids the persistent build-up of error that global fusion causes. Guided by those correspondences, Geometric Correspondence Attention selectively injects geometrically matched historical features into noisy target patches during denoising.

Occlusion needs its own handling, because a correspondence can make geometric sense and still point at something hidden. An Invisible Octree accumulates visibility evidence and rejects correspondences that are geometrically plausible but occluded.

Extensive experiments report state-of-the-art visual quality, precise camera control, and revisit consistency, so a place should look the same when the camera comes back to it. Together these enable minute-long video along challenging trajectories. The abstract gives no numerical margins.

The design choice is to ask less of geometry. Rather than trusting a reconstructed world to be right, GEAR uses geometry only as a pointer and leaves the recovery of content to attention, so any error in the pointer stays local to one lookup instead of being baked into a shared memory.