Memory-V2V：为视频到视频扩散模型增强记忆

Recent foundational video-to-video diffusion models have achieved impressive results in editing user-provided videos by modifying appearance, motion, or camera movement. However, real-world video editing is often an iterative process, where users refine results across multiple rounds of interaction. In this multi-turn setting, current video editors struggle to maintain cross-consistency across sequential edits. Models often lose stylistic, content, or structural information established in previous editing rounds when handling multiple turns, leading to results generated in subsequent edits that are uncoordinated with earlier versions, thereby degrading user experience and editing efficiency. For instance, after performing multiple style transfers or object replacements on a video, the model might fail to maintain object identity consistency across different rounds, or the overall video style might drift during continuous edits. The key to addressing this challenge lies in introducing a memory mechanism that allows the model to store and retrieve critical information across editing rounds. This memory mechanism could involve encoding user intent, editing history, and intermediate representations of video content at different stages of editing. By integrating a memory module, the model can refer to previous editing states when performing new edits, thereby ensuring high consistency and coherence throughout the entire video editing process. This not only improves the quality of the editing results but also significantly enhances user satisfaction in iterative editing workflows.