Self-learning style transfer via dual-path feature fusion in diffusion models

Style transfer combines the content of one image with the artistic style of another. Recent advances in diffusion models have improved image generation, but reducing style-content coupling while maintaining high-quality generation remains challenging. We present Self-Learning Style Transfer (SLST), a diffusion-based framework that coordinates style-dominant and content-oriented representations through dual-path feature fusion. For style representation, the Self-Learning Style Encoder (SLS-Encoder) uses learnable queries, random masking, and cross-view consistency learning to encourage style-dominant features with reduced content dependence, without requiring paired content-style data. For content preservation, adaptive edge extraction provides structural cues, while BLIP-2 generates concise content-focused descriptions for semantic guidance. Quantitative evaluations, visual comparisons, and user studies indicate favorable style-transfer behavior in the evaluated settings. These downstream results suggest reduced content interference but do not directly establish representation-level disentanglement.