A research team led by Professor Lee JaeKoo of the School of Artificial Intelligence at Kookmin University (President Jeong Seung Ryul) is leading the way in global AI technology by presenting “FEAT,” a next-generation virtual fitting and fashion editing technology, as a highlight at “CVPR (Computer Vision and Pattern Recognition) 2026,” the world’s most prestigious conference on computer vision and artificial intelligence.

Under the guidance of Professor Lee JaeKoo (corresponding author) at Kookmin University, Kwon So-ye, a master’s student in the Department of Computer Science (first author), played a pivotal role in leading the research. The paper, titled “FEAT: Fashion Editing and Try-On from Any Design,” presents an innovative methodology that overcomes the limitations of existing technologies—which relied solely on specific clothing images—by enabling the use of any non-apparel design source, such as abstract artworks or general photographs, as fashion items. This achievement is drawing exceptional attention from the academic community, particularly because it was led by a master’s student with relatively limited research experience serving as the first author. This is because CVPR Highlight papers—awarded only to a select few—are typically led by doctoral-level researchers from leading global tech companies and prestigious universities.

■ Kookmin University Research Team’s Innovative Idea Solves Practical Challenges in AI Application

At the heart of this achievement are two key algorithms that address practical challenges in the fashion industry.

Existing generative AI models suffered from a critical flaw known as “content leakage,” where unnecessary elements (e.g., the face or background of a person in the original photo) were synthesized onto the clothing when extracting designs from images. To overcome this, the Kookmin University research team devised the ‘DDI (Disentangled Dual Injection)’ technology, successfully separating the design’s content and style completely to precisely inject only the textures and colors desired by the user.

Additionally, to address the issue of unnatural traces left by the original clothing when dressing a model in new clothes, the team introduced the ‘OGNF (Orthogonal-Guided Noise Fusion)’ technique. This groundbreaking method effectively suppresses the areas of the existing clothing using orthogonal projection without the need for additional large-scale retraining. This enables perfect virtual try-ons not only for tops and bottoms but also for various accessories—such as bags, shoes, scarves, and belts—which previous models struggled to handle.

■ “Kookmin University’s Power to Lead the Fashion Industry’s Digital/AI Transformation”

The FEAT technology led by Kookmin University is expected to have a massive ripple effect across the entire fashion industry, going beyond a mere academic achievement. By allowing consumers to preview the entire outfit, including accessories, it increases customer satisfaction and, furthermore, can dramatically reduce return rates—the primary cause of cost losses in online shopping. Furthermore, by replacing the production of physical samples during the planning stage with AI, the technology paves the way for the realization of “Sustainable Fashion” by preventing fabric waste, and has opened up the possibility for K-AI technology to shift the paradigm of the global fashion tech market.

Professor Lee JaeKoo of Kookmin University stated, “This achievement is a source of great pride, resulting from a Kookmin University master’s student’s persistent optimization of the model and their confident competition with world-class researchers.” He added, “The students’ exceptional dedication and original ideas were the key drivers behind the leap forward in AI virtual fitting technology, and we will continue to provide unwavering support so that Kookmin University can continue to produce key talent in the global fashion tech and artificial intelligence fields.” This paper and its results, which demonstrate Kookmin University’s outstanding AI research capabilities, have been made publicly available along with the source code and virtual fitting dataset to facilitate technological advancement and reproducibility. They were officially presented to researchers worldwide at the CVPR 2026 conference.