<p>This study focuses on the intelligent rendering task of “mirror painting”, aiming to develop a high-quality, real-time rendering method applicable to digital art, human-computer interaction, and game art scenarios. The core challenge lies in achieving high-fidelity local texture modeling and global style consistency under limited computational resources. Existing methods such as StyleSwin, Style-Aware Network (SANet), and Stable Diffusion exhibit weak structural consistency, severe style drift, or high computational costs, making it difficult to simultaneously balance speed and image quality. To address these challenges, this study proposes an intelligent rendering method called the Mamba-Swin-Low-Rank Adaptation (LoRA) rendering framework. The method integrates three key technologies. First, local features are extracted based on the Mamba state space model to effectively preserve brushstrokes and edge details. Second, Swin Transformer for global feature modeling is introduced to reduce complexity through sliding window attention mechanisms. Finally, style transfer optimization is carried out in combination with LoRA, maintaining style consistency while updating only a small number of parameters. Experimental results demonstrate that the method achieves 34.2 dB in Peak Signal-to-Noise Ratio, 0.91 in Structural Similarity Index Measure, and 0.18 in Learned Perceptual Image Patch Similarity for image quality assessment. The proposed model outperforms current mainstream methods in image quality, computational efficiency, and interactive experience, showing strong potential for practical applications.</p>