GEAR octree memory enables minute-long video consistency
GEAR uses octree attention as constant-size memory, generating consistent videos up to 60 seconds with precise camera control. The approach replaces global 3D fusion with token-level geometry addressing, but currently handles only static scenes.
- GEAR generates consistent videos up to 60 seconds long
- Constant-size memory replaces growing global 3D fusion
- Method does not cover dynamic objects or occlusions
- Invisible octree implementation needs extra work for other codebases
Read next
Science