LoRA 기반 파인튜닝을 활용한 생성형 AI 패션 디자인 연구- 브랜드 RE;CODE의 디자인 특성 재현 및 확장 - = Generative AI Fashion Design with LoRA Fine-tuning- Reproducing and Extending the Design Identity of RE;CODE -

다국어 초록 (Multilingual Abstract)

This study proposes a LoRA-based fine-tuning methodology that enables creative product-line expansion while preserving a fashion brand’s design DNA. It addresses the limitation of general-purpose image generation models in consistently reproducing brand identity and empirically tests whether core design DNA can be transferred to new product categories. We constructed a dataset of 8,465 product images from the upcycling brand RE;CODE and trained SDXL-based LoRA models for five product categories. To better capture brand philosophy and design attributes, we developed an in-house intelligent captioning pipeline using Gemma 3, which integrates visual features and brand metadata to generate captions for optimizing SDXL’s CLIP text encoder. Using a Cross-Generation setup, we quantitatively examined design DNA transfer when category-specific LoRA models generate unseen categories, via FID, CLIP Score, and LPIPS. All LoRA models improved FID over the base SDXL model by 17.5%–40.9%, indicating strong brand-style fidelity. Notably, the outerwear LoRA achieved a 37.4% FID improvement when generating dresses, confirming successful transfer of structural and aesthetic characteristics. CLIP Scores remained stable across experiments, suggesting a balance between prompt fidelity and creativity, while increased LPIPS indicates effective rendering of RE;CODE’s complex material mixtures and irregular textures. These findings demonstrate that LoRA fine-tuning can function as a creative partner beyond mere style imitation, thereby supporting scalable brand-consistent generation. The proposed method offers an efficient and accessible approach for individual designers and small-to-medium brands, contributing to digital transformation and sustainable design processes in fashion.