ColorFlow implements a three-stage retrieval-augmented pipeline for colorizing image sequences such as cartoons or comics while preserving character and object identity across frames. The approach uses a dual-branch design with separate branches for color identity extraction and colorization, leveraging diffusion models’ self-attention mechanism for in-context learning and color matching. Unlike previous methods that require per-identity fine-tuning, this framework provides a robust and generalizable solution suitable for industrial applications.