Flow Residual Segmentation and Generative Reconstruction for Motion-Aware Video Coding

Abstract

Recent diffusion-based video generation methods have shown strong potential for video compression and frame interpolation; however, they often suffer from unstable motion modeling and excessive computational overhead. To address these issues, we propose a Flow Residual Segmentation (FRS) algorithm that adaptively divides a video sequence into segments where optical flow remains consistent, ensuring that motion within each segment can be reliably reconstructed. Building upon this segmentation, we develop a generative video compression framework that integrates the FRS-based segment selection with a diffusion-driven interpolation model. The proposed method achieves efficient, flow-consistent reconstruction even at extremely low bitrates by balancing motion complexity and keyframe allocation. Extensive experiments demonstrate that, compared with Extreme Video Compression and recent generative interpolation approaches, our method achieves superior perceptual quality (lower LPIPS), stable temporal consistency, and improved efficiency under real-world compression scenarios.

ZIYUE ZENG

Ph.D. Student at Waseda University

Pursuing socially meaningful, grounded research with enduring passion and diligence.