Earticle

자전거 이용 증가와 함께 자전거도로의 사고 예방 및 위험 상황에 대한 영상 기반 모니터링 의 중요성이 더욱 부각되고 있다. 그러나 계절·조도·기상 변화가 충분히 반영된 영상 데이터를 확보하는 데에는 많은 시간이 소요되고, 설상 가상으로 라벨링 비용 또한 높아 객체 탐지 모델 개발에 제약이 발생한다. 본 연구는 이러한 문제를 해결하기 위해 구조적 제약과 스타일 적응 을 동시에 반영하는 영상 합성 기법을 제안하였다. 제안한 방법은 Stable Diffusion 기반 모델에 ControlNet과 Low-Rank Adaptation을 결합하여, 마스크 영상을 통한 구조 제어와 스타일 미세조 정을 통합적으로 수행한다. 실제 CCTV 영상을 기반으로 데이터세트를 구축하고, 세 가지 Stable Diffusion 계열 기저 모델을 대상으로 합성 성능을 비교하였다. 성능 평가는 Fréchet Inception Distance와 CLIP-score를 활용하였으며, 그 결과 제안한 방법이 사실성과 텍스트 정합 성 측면에서 우수한 합성 품질을 달성함을 확인하였다. 또한 텍스트 프롬프트 조작만으로 계 절 및 기상 조건을 반영한 영상 생성이 가능함을 검증하였다. 본 연구는 촬영이 어려운 다양한 환경 조건의 데이터를 효율적으로 생성할 수 있어 자전거도로 모니터링을 위한 데이터 부족 문제 해결에 기여하며, 향후 객체 탐지 및 안전관리 기술의 고도화에 효과적으로 활용될 수 있다.

The growing use of bicycles has heightened the importance of video-based monitoring for preventing accidents and detecting hazardous situations on bicycle roads. However, collecting video data that adequately reflects variations in season, illumination, and weather requires substantial time, and the high cost of data labeling further limits the development of effective object-detection models. To address these challenges, this study proposes an image synthesis method that simultaneously incorporates structural constraints and style adaptation. The proposed approach integrates Stable Diffusion with ControlNet and Low-Rank Adaptation (LoRA), enabling unified control of scene structure through mask images and fine-grained style adjustment. A dataset was constructed using real CCTV footage, and three Stable Diffusion–based backbone models were evaluated for their synthesis performance. Fréchet Inception Distance and CLIP-score were used for quantitative assessment, demonstrating that the proposed method achieves superior realism and semantic alignment between images and text. Furthermore, the model successfully generated images reflecting seasonal and weather variations solely through prompt manipulation. This research provides an efficient solution for generating diverse environmental conditions that are difficult to capture in practice, thereby alleviating data scarcity in bicycle-road monitoring and supporting the advancement of nextgeneration object-detection and safety-management technologies.

The widespread discussion of AI-generated content, fueled by the emergence of consumer applications like ChatGPT and Midjourney, has attracted significant attention. Among various AI applications, AI painting has gained popularity due to its mature technology, user-friendly nature, and excellent output quality, resulting in a rapid growth in user numbers. Midjourney and Stable Diffusion are two of the most widely used AI painting tools by users. In this study, the author adopts a perspective that represents the general public and utilizes case studies and comparative analysis to summarize the distinctive features and differences between Midjourney and Stable Diffusion in the context of AI character illustration. The aim is to provide informative material for those interested in AI painting and lay a solid foundation for further in-depth research on AI-generated content. The research findings indicate that both software can generate excellent character images but with distinct features.

With significant advancements in the field of artificial intelligence, many novel algorithms and technologies have emerged. Currently, AI painting can generate high-quality images based on textual descriptions. However, it is often challenging to control details when generating images, even with complex textual inputs. Therefore, there is a need to implement additional control mechanisms beyond textual descriptions. Based on ControlNet, this passage describes a combined utilization of various local controls (such as edge maps and depth maps) and global control within a single model. It provides a comprehensive exposition of the fundamental concepts of ControlNet, elucidating its theoretical foundation and relevant technological features. Furthermore, combining methods and applications, understanding the technical characteristics involves analyzing distinct advantages and image differences. This further explores insights into the development of image generation patterns.

화장품 립 제품 개발 현장에서 배합비 변경 시마다 시제품 재제조와 육안 확인을 반복해야 하는 비효율을 화장품 립 제품의 성분 배합비(유기 색소, 무기 안료, 왁스, 오일, 진주 광택제, 필름 형성제)를 정량적 입력값으로 활용해, 발색 색상과 질감 특성을 시각적으로 시뮬레이션하는 인공지능 기반 립 메이크업 생성 시스템을 제안한다. 기존의 머신러닝 기반 연구는 수치적 물성 예측에 그쳐 시각적 확인이 어려웠고, GAN 및 확산 모델 기반의 가상 메이크업 연구는 실제 화장품의 정량적 배합 정보를 반영하지 못한다는 한계가 있었다. 이러한 문제를 해결하기 위해 성분 배합비를 정규화된 벡터로 변환하고, 다층 퍼셉트론(MLP)을 통해 색상, 광택, 불투명도 등의 시각적 속성을 예측하는 구조를 확립하였다. 나아가 예측된 색상 데이터를 공간적 조건 지도로 변환하여 ControlNet 의 입력으로 활용함으로써 배합비에 부합하는 정밀한 색상을 구현하였다. 또한, 원료 특성 기반의 유효 광택 지수를 정의하고, 이에 따라 생성 모델의 컨디셔닝 스케일과 디노이징 강도를 동적으로 제어함으로써 입술의 질감 표현력을 극대화하였다. 정량적·정성적 실험 및 화장품 업계 전문가들을 대상으로 한 사용성 평가를 통해, 본 시스템이 실제 R&D 현장에서 시제품 제작 루프를 단축하고 초기 배합 설계의 예측력을 높이는 데 실질적으로 기여할 수 있음을 입증하였다.

In the development process of cosmetic lip products, repeated prototype manufacturing and visual inspection are often required whenever the formulation ratio changes, resulting in significant inefficiency. To address this issue, this paper proposes an artificial intelligence-based lip makeup generation system that visually simulates lip color and texture characteristics using quantitative cosmetic formulation information, including organic pigments, inorganic pigments, waxes, oils, pearlescent agents, and film formers. Existing machine learning-based approaches have mainly focused on numerical property prediction, making visual verification difficult, while GAN- and diffusion-based virtual makeup studies have limitations in reflecting quantitative cosmetic formulation information. To overcome these limitations, formulation ratios were converted into normalized vectors, and a multilayer perceptron (MLP)-based regression structure was designed to transform formulation information into visual attributes such as color, glossiness, and opacity. Furthermore, the predicted color information was converted into a spatial condition map and utilized as an input condition for ControlNet, enabling generation results that reflect formulation-based color characteristics. In addition, an effective gloss index based on ingredient characteristics was defined, and the conditioning scale and denoising strength of the generation model were dynamically adjusted according to the gloss characteristics to improve lip texture representation. Through quantitative and qualitative experiments, as well as a usability evaluation involving cosmetic industry practitioners, the proposed system demonstrated the potential to reduce repetitive prototype production processes and support early-stage formulation design in cosmetic R&D workflows.