[论文] Consistent Video-to-Video Translation via Explicit Correspondences
研究领域: CV 作者: Gaurav Parmar, Zhengqi Li, Richard Zhang 发布时间: 2025-11-13 arXiv: 2511.10591
论文概要
研究领域: CV 作者: Gaurav Parmar, Zhengqi Li, Richard Zhang 发布时间: 2025-11-13 arXiv: 2511.10591
中文摘要
交互式视频到视频应用需要实时生成并保持长程时间一致性。近期方法将注意力限制在固定长度的因果窗口内,限制了长程一致性。我们提出 vid2vid-long,基于一个简单观察:输入视频中匹配的区域在输出中也应匹配。具体而言,在输入中计算块级对应关系,引导检索先前生成输出的对应区域块,并通过对应交叉注意力作为模型条件,在保持生成质量与实时性能的同时持续提升长程时间一致性。
原文摘要
Interactive video-to-video applications require real-time generation while maintaining long-range temporal consistency. Recent methods restrict attention to a fixed-length causal window, limiting long-range consistency. We introduce vid2vid-long, a correspondence-based approach grounded in the observation that matching regions in the input video should match in the output. We compute patch-level correspondences in the input, guide retrieval of previously generated outputs, and condition the model via correspondence cross-attention, consistently improving long-range temporal consistency while preserving quality and real-time performance.
*自动采集于 2026-11-15*
#论文 #arXiv #CV #小凯