A Diffusion-Based Approach for Mogu Painting Style Transfer

Recent advancements in deep learning have propelled style transfer as a critical technique in computer vision, enabling the transformation of artistic styles across diverse images. While considerable progress has been made in areas such as anime, impressionism, and modern art, the domain of traditional Chinese painting remains relatively underexplored. Among its unique styles, the Mogu (Boneless) technique, pioneered by artists like Yun Shouping, is characterized by fluid brushstrokes and subtle gradient blending, resulting in ethereal and dynamic depictions. Motivated by this distinctive artistic process, we introduce a novel neural network framework for Mogu-style transfer. Our approach begins with the extraction and optimization of edge features, followed by the fusion of content and edge features to enhance the representation of floral subjects. We then incorporate Yun Shouping’s signature color techniques and blend style and content through a U-Net-based diffusion model. Finally, global optimization is applied to refine the entire image, ensuring the generated output closely adheres to the stylistic principles of Mogu painting. To validate the effectiveness of our framework, we construct a dataset of Mogu paintings by Yun Shouping and conduct comprehensive experiments. The results highlight the framework’s ability to produce visually compelling and stylistically faithful outputs.