ColorFM: An Optimization-to-Learning Framework for Color Transfer via Flow Matching

Overview

ColorFM addresses the long-standing trade-off in color transfer between the high precision of iterative, instance-specific optimization and the speed of feed-forward neural networks. By reformulating color transfer as the transport of pixel distributions along velocity fields via Flow Matching (FM), the authors create a unified framework that bridges these two paradigms.

The framework consists of two components: ColorFM-O and ColorFM-L.

• ColorFM-O (Optimization): This component acts as an offline generator. It uses hierarchical color coupling and semantic priors to optimize a velocity field for specific image pairs. By numerically integrating the resulting flow trajectories, it produces high-fidelity, semantically consistent stylized images that serve as pseudo-ground truth for training.
• ColorFM-L (Learning): This is the efficient, feed-forward inference model. It uses implicit state modeling to predict flow parameters based on content and style features. By leveraging the geometric regularity of the trajectories learned by ColorFM-O, it performs color transfer via bidirectional linearized transport, enabling rapid, one-step inference that maintains the quality of the optimization-based approach.

Existing color transfer methods often struggle with either computational intensity (optimization-based) or poor generalization and visual artifacts (learning-based). ColorFM effectively combines the strengths of both, providing a robust solution that preserves structural fidelity and semantic consistency while enabling real-time performance. This makes it a highly practical tool for high-quality, automated image retouching.