Emotional Stylization Based on Diffusion Models

In recent years, deep learning based image generation has progressed beyond traditional pixel-level reconstruction toward semantically controllable content synthesis. In emotional stylization, diffusion models, with their progressive denoising mechanism and high-fidelity generation capability, have introduced new possibilities for emotion-aware image stylization. However, emotion, as a high-level semantic concept, exhibits subjectivity and ambiguity when mapped to visual stylistic attributes such as color tone and texture patterns. Consequently, existing methods face challenges in achieving precise emotional expressiveness and style controllability. To address these issues, this paper proposes an emotional stylization framework based on Stable Diffusion, integrating the ControlNet architecture with a Low-Rank Adaptation (LoRA) fine-tuning strategy. Specifically, the UNet module within the diffusion model is directionally optimized using the ArtEmis style-emotion dataset to learn the correspondence between emotional semantics and artistic styles through color dynamics and texture features. Experimental results demonstrate that the proposed method effectively enhances emotional expressiveness, style controllability, and semantic consistency, outperforming existing approaches in emotional stylization.