CloZ: Natural Language Guided Clothing Design System

CloZ is a clothing design system facilitated by natural language prompts. Cloz supports two main functions; 1) generating clothing images via natural language prompting. 2) editing generated images by replacing keywords from previous prompts.
Inspired by FACAD [1], we first built the nordstrom96568 dataset, which consists of 96568 (prompt, clothing image) pairs. Then we trained stable diffusion [2] 2.1 with our dataset to generate clothing images using prompts. The editing function was implemented by CycleDiffusion [3]. Also, we designed the CloZ's web-based interface based on guidelines of prior research [4].
AI Engineering
데이터 세트 수집 및 전처리
모델 파이프라인 설계
Cycle Diffusion을 이용한 Latent space 조작 및 실험
LoRA를 이용한 Stable Diffusion 모델 PEFT
FastAPI 서빙 서버 개발
A100 인스턴스 상에서 모델 서빙
T4 * 4 AWS Instance 상에서 멀티 지피유 로드밸런싱
Frontend Engineering
Styled-component 기반 시스템 개발