Ada-VE: Training-Free Consistent Video Editing Using Adaptive Motion Prior (WACV 2025)
Tanvir Mahmud, Mustafa Munir, Radu Marculescu, and Diana Marculescu
We propose an adaptive motion-guided cross-frame attention mechanism for video synthesis that reduces redundant computations by focusing on moving regions and reuses key-values across frames, achieving smoother transitions, enhanced quality, and processing more frames within the same computational budget. This approach outperforms current methods in accuracy and temporal consistency.
Code will be released soon!