GEAR's octree memory enables minute-long consistent video generation
Researchers developed GEAR, a new video generation method that maintains visual consistency for up to 60 seconds. Traditional approaches either used slow memory lookups or accumulated geometric errors in global 3D representations. GEAR instead builds an 'invisible' octree memory structure that routes attention through constant-size memory during denoising. This allows precise camera control and revisit consistency on challenging trajectories without increasing memory requirements. The paper focuses on static scenes, leaving dynamic object handling and full pipeline integration as open questions.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in