Towards Coherent Video Colorization: When Optical Flow Meets Image Diffusion Models

Abstract

Existing diffusion video colorization suffers from costly training and temporal flickering. We propose an optical flow-guided diffusion model with cross-frame attention for consistent color propagation under complex motions, plus reference image guidance for higher color fidelity. A hybrid synthetic data pipeline lowers real data reliance via two-stage training. Evaluations validate our advantages in training efficiency, temporal coherence and color quality.