Geometry as Address: Routing Attention to Visual Memory for Long-Horizon Camera-Controlled Video Generation
GEAR uses geometry to route attention to visual memory for long camera-controlled video generation.
GEAR is a Geometry-Enabled Attention Routing method for long-horizon camera-controlled video generation. It keeps past frames as latents and uses per-frame geometry only to address which visual memory tokens to read, avoiding error buildup from a persistent global 3D reconstruction. Geometric Correspondence Attention injects matched historical features into noisy target patches during denoising, while an Invisible Octree filters geometrically plausible but occluded matches. Experiments report state-of-the-art visual quality, precise camera control, revisit consistency, and minute-long generation on difficult trajectories.
- GEAR routes attention using geometry as a memory address.
- History stays as frame latents rather than a fused 3D scene.
- Geometric Correspondence Attention injects matched features during denoising.
- An Invisible Octree rejects occluded correspondences.
- Authors report minute-long, camera-controlled video with revisit consistency.
From the original article
1 min read at huggingface.co
Long-horizon camera-controlled video generation requires recovering previously observed content from an ever-growing visual history. Existing approaches either search historical context implicitly or reconstruct it into persistent 3D memory, facing inefficient memory access or accumulated geometric errors.
By Zesong Yang, Weikai Chen, Liyuan Cui · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.