GEAR framework routes attention using geometric addresses
September 27, 2026
GEAR (Geometry-Enabled Attention Routing) enables long-horizon video generation by using per-frame geometry as an explicit token-level address for visual memory. This avoids the memory inefficiency of 3D reconstruction by using attention to recover specific historical frame latents.
HOW THIS AFFECTS YOU
●
builderYou can achieve more efficient long-form video generation by leveraging geometric routing instead of dense 3D memory.
●
researcherThis provides a new method for combining explicit geometric priors with implicit attention mechanisms.