AI
Geometry as Address: Routing Attention to Visual Memory for Long-Horizon Camera-Controlled Video Generation
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 27/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 04:02:28 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 16:18:23 UTC.
Briefings
This research introduces GEAR, a framework for long-horizon video generation that uses geometry as an explicit token-level address for visual memory. Unlike prior methods that fuse historical observations into persistent 3D representations, GEAR retains frames as latents and relies on per-frame geometry to establish correspondences with target views. This approach avoids error accumulation from global fusion, enabling precise camera control and minute-long video generation along challenging trajectories. The system employs Geometric Correspondence Attention (GCA) to inject geometrically matched features into noisy patches and uses an Invisible Octree to filter out occluded correspondences.
Key points
- GEAR retains historical frames as frame latents rather than fusing them into a persistent global 3D representation to avoid error accumulation.
- Per-frame geometry establishes token-level correspondences with target views, while attention selects which historical content to recover.
Context and limitations
The framework relies on per-frame geometry for correspondences and an Invisible Octree for visibility evidence, which may introduce geometric inaccuracies in occluded regions. Long-horizon generation along challenging trajectories remains a novel capability not yet fully quantified.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 04:02:28 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigddkbfn6o24x6ymy6ig67vv4dfvfnf4elzujlbeaoqx4zud5kxqu