Unified Video Editing
A single editor handles diverse text-guided video editing cases in an in-context formulation.
Abstract
LIVEditor targets efficient unified video editing. Instead of treating long video context with dense attention, it introduces In-Context Sparse Attention to retrieve and attend to the most relevant contextual tokens. The resulting editor keeps the in-context editing formulation while substantially reducing attention cost, making high-quality video editing faster and easier to deploy.
Highlights
A compact summary of the method and demo focus.
In-Context Sparse Attention
ISA prunes redundant context and routes queries to relevant source-video tokens.
Fast Inference
The demo focuses on visual comparison results while the paper reports large attention-latency reductions.
Gallery
Representative visual comparisons for full-attention editing and ISA-accelerated editing.
Comparison Results
We compare source videos, full-attention editing, and ISA-accelerated editing results.
Full-Attention vs. ISA Comparisons
Each row is a side-by-side comparison of Source, Full Attention, and ISA under the same editing instruction.
SourceFull AttentionISA
SourceFull AttentionISA
SourceFull AttentionISA
SourceFull AttentionISA
SourceFull AttentionISA
SourceFull AttentionISA
Source vs. ISA Editing Results
Each pair shows the source/reference video and the edited result produced by ISA acceleration.
Source
ISA Result
Source
ISA Result
Source
ISA Result
Source
ISA Result
Source
ISA Result
Source
ISA Result
Source
ISA Result
Source
ISA Result
Source
ISA Result