The paper introduces RB-Modulation, a novel training-free method for personalizing diffusion models. RB-Modulation addresses the challenges of style extraction, content leakage, and effective composition of style and content from a single reference image. The method is built on a stochastic optimal controller where a style descriptor encodes desired attributes through a terminal cost. This ensures high fidelity to the reference style and adherence to the given text prompt. The paper also introduces an Attention Feature Aggregation (AFA) module that decouples content and style from the reference image, enhancing image fidelity. The effectiveness of RB-Modulation is demonstrated through extensive experiments in stylization and content-style composition, showing superior performance over state-of-the-art methods in terms of human preference and prompt alignment metrics. The theoretical justifications and practical implementation details are provided, along with a user study validating the method's effectiveness.The paper introduces RB-Modulation, a novel training-free method for personalizing diffusion models. RB-Modulation addresses the challenges of style extraction, content leakage, and effective composition of style and content from a single reference image. The method is built on a stochastic optimal controller where a style descriptor encodes desired attributes through a terminal cost. This ensures high fidelity to the reference style and adherence to the given text prompt. The paper also introduces an Attention Feature Aggregation (AFA) module that decouples content and style from the reference image, enhancing image fidelity. The effectiveness of RB-Modulation is demonstrated through extensive experiments in stylization and content-style composition, showing superior performance over state-of-the-art methods in terms of human preference and prompt alignment metrics. The theoretical justifications and practical implementation details are provided, along with a user study validating the method's effectiveness.