<![CDATA[The rapid advancement of Generative Artificial Intelligence has accelerated the adoption of diffusion models in fashion design applications. However, conventional text-to-image approaches often encounter limitations in maintaining visual consistency and controllability during image generation. This study proposes a multimodal Muslimwear design generation system based on Stable Diffusion v1.5 by integrating textual prompts and reference images through a cross-attention fusion mechanism. The training dataset combines DeepFashion1 and a curated Muslimwear dataset, which were preprocessed through image normalization, resolution standardization, and automated caption generation using BLIP. Domain adaptation was performed using the Low-Rank Adaptation (LoRA) technique to enable computationally efficient fine-tuning. Performance evaluation employed Fréchet Inception Distance (FID) and Structural Similarity Index Measure (SSIM) to assess visual quality and structural consistency. Experimental results indicate that the female model achieved a FID score of 176.77 and an SSIM score of 0.311, outperforming the male model with a FID score of 256.22 and an SSIM score of 0.275. The findings demonstrate that multimodal conditioning enhances visual distribution learning and structural preservation, supporting the development of controllable and efficient AI-assisted fashion design systems.]]>
Copyrights © 2026
