Optimization Path for Chinese Portrait Generation with AIGC via SAM Segmentation and LoRA Fine-Tuning

With the rapid development of generative artificial intelligence, the current AIGC portrait generation faces issues in character images such as a high proportion of Western cultural elements, weakened representation of local culture, abuse of cultural elements, and semantic deviation from underlying connotations, which makes it difficult to meet the demand for accurate expression of local cultural characteristics. This paper proposes to use the precise image segmentation technology of the Segment Anything Model (SAM) to extract core cultural elements from Chinese portraits and build a high-quality database, laying a foundation for the semantic analysis and structural reconstruction of cultural codes. By classifying and summarizing cultural codes, the Low-Rank Adaptation (LoRA) model is adopted to perform lightweight fine-tuning on existing open-source multimodal AI models. While retaining the generative capability of the pre-trained model, targeted correction is conducted for the semantic deviations of Chinese portraits, ensuring that the generated portraits not only restore the detailed features of Chinese portraits but also conform to their cultural connotations. Compared with current AIGC portrait generation methods, the scheme proposed in this paper achieves a balance between technical generation and cultural inheritance through precise image segmentation and targeted model fine-tuning, providing a feasible paradigm for the preservation and innovation of Chinese portraits in the digital era.