Specific Style Image Generation Based on LoRA Fine-Tuning Technology

This paper aims to improve the performance of general Text-to-Image models for image generation in a specific art style.While foundational models generate diverse imagery, they often fail to capture the unique nuances required for specialized artistic genres and instead produce generic results.To address the high computational cost of full fine-tuning, an efficient method based on Low-Rank Adaptation is proposed.This study enables the pre-trained Stable Diffusion v1.5 model to learn and reproduce the Japanese Ukiyo-e style, characterized by its distinct flat colors and bold outlines.A small, high-quality dataset of Ukiyo-e artworks was constructed and meticulously annotated, followed by parameterefficient fine-tuning using low-rank adaptation (LoRA).Both qualitative and quantitative evaluation schemes were implemented to validate the method.Experimental results show that the method consistently generates images in the target art style while maintaining semantic consistency with the input text.This approach provides a practical reference for the application of AIGC to personalized and professional image generation.