Color Consistency Anime Style Transfer Based on Hybrid Structural Decomposition Network

In the anime industry, character design frequently encounters challenges such as oversimplification and homogenization, which severely restrict the diversity and creativity of character ap-pearances. To tackle these issues, we introduce a Hybrid Structural Decomposition Network (HSDN) that aims to generate anime characters with a realistic style by learning the expressive techniques of traditional anime, thereby providing valuable inspiration and references for designers. The proposed HSDN utilizes an encoder to separately extract structural and color features from input images. The structural features are then processed by our style transfer sub-network, which is based on diffusion models. Subsequently, the structural and color features are fused at multiple scales, and a specially designed decoder is employed to generate the final style-transferred image. Additionally, we have incorporated specific skip connections to mitigate local detail loss during image generation and to enhance the stability of the diffusion process. Experimental results demonstrate that our proposed model not only visually generates anime characters with realistic styles but also achieves superior performance in terms of Style FID and Semantic FID compared with state-of-the-art methods.