VideoGrain
VideoGrain is an official implementation of the ICLR 2025 paper "Modulating Space-Time Attention for Multi-Grained Video Editing." It is a zero-shot, training-free method designed for advanced video manipulation without requiring model fine-tuning. The software enables multi-grained editing capabilities, including class-level editing for objects within the same category, instance-level editing to transform individual instances into distinct objects, and part-level editing for modifying specific attributes or adding new elements to existing objects. A core feature of VideoGrain is its modulation of cross-attention and self-attention mechanisms. By adjusting cross-attention, it supports multi-region control using a single prompt, while self-attention modulation facilitates feature decoupling through clustering. The approach addresses limitations in previous state-of-the-art methods that were restricted to class-level editing. VideoGrain is suitable for researchers and developers needing flexible video editing t