PermaVid: Consistent Video Generation Across Edits via Disentangled Context Memory
Consistent video generation across global and local edits
Abstract
PermaVid preserves consistency across video edits, camera viewpoints, and time by separating context memory into RGB appearance and depth geometry. An edit-aware update and retrieval strategy refreshes modified content while retaining reusable scene structure. A memory-guided video generator combines references from both memory banks to maintain coherent scenes after global style changes and local object edits.
Type
Publication
In Advances in Neural Information Processing Systems (NeurIPS)
Tong WU 吴桐
Assistant Professor @ Fudan
My research interests include 3d vision, long-tailed recognition, and robustness.