Consistent Video-to-Video Translation via Explicit Correspondences
vid2vid-long reuses matching regions from earlier outputs to keep generated video consistent beyond a model’s context window.
Fast video-to-video models often restrict attention to a fixed stretch of recent frames, losing relevant history. The proposed method finds matching patches across input frames and retrieves the corresponding patches from previously generated outputs. It feeds those patches into the model through correspondence cross-attention. The authors report improved long-range temporal consistency across recent video models while preserving generation quality and real-time performance.
ArXiv · AI/CL/LG's note
Fast video-to-video models often restrict attention to a fixed stretch of recent frames, losing relevant history. The proposed method finds matching patches across input frames and retrieves the corresponding patches from previously generated outputs. It feeds those patches into the model through correspondence cross-attention. The authors report improved long-range temporal consistency across recent video models while preserving generation quality and real-time performance.
ArXiv · AI/CL/LG's note
score 5