A paper titled “Nonlinear Color Transfer via Learnable Bezier Flows,” authored by a research team led by Professor Kim Jangho of the Department of Artificial Intelligence at Kookmin University in collaboration with EziWeed, was presented at CVPR 2026, the premier international conference in the field of computer vision.

CVPR (IEEE/CVF Conference on Computer Vision and Pattern Recognition) is a world-renowned academic conference representing the fields of computer vision and pattern recognition. It is regarded as one of the most prestigious international conferences where various visual intelligence research topics—including AI-based image understanding, image generation, 3D vision, medical imaging, autonomous driving, and robotics—are presented. Every year, researchers from universities, research institutions, and global companies around the world participate in this event, which has established itself as a platform for sharing the latest computer vision technologies and their potential industrial applications.

This research was a joint effort between the Eziwid research team and Professor Kim Jangho from the Department of Artificial Intelligence at Kookmin University. The paper addresses a color transfer technique that naturally reflects the color tones and atmosphere of a target image while preserving the structure and texture of the source image.

The research team noted that existing color transfer techniques can cause unnatural color bleeding or structural distortion in images with complex lighting, textures, or 3D rendering. To address this, they proposed NCT (Nonlinear Color Transfer), a learnable, Bezier Flow-based nonlinear color transfer method. A key feature of NCT is that it is designed so that colors do not simply follow a linear path but move naturally along nonlinear paths tailored to the characteristics of the image.

Additionally, to accommodate diverse color distributions and lighting conditions, the team applied a Mixture of Experts (MoE) architecture, enabling more stable transformations based on the color characteristics of each image. Experimental results showed that the proposed method outperformed existing color transformation techniques in terms of content structure preservation and visual naturalness, and its high applicability was confirmed even with 3D rendering-based media art data.

This achievement is significant in that it represents a collaboration between academia and industry to develop AI-based image processing technology that can be utilized in real-world content production environments. In particular, it is expected to reduce color retouching costs and support more natural color conversion in areas such as high-resolution media art, 3D rendering content, digital exhibitions, and advertising and brand video production.

Eziwid and the research team led by Professor Kim Jangho of Kookmin University plan to expand their research beyond static images to include video content in the future, continuing to develop temporally consistent, high-quality color conversion technology and conducting various generative AI studies.

△ Lee Jun-hyung, Researcher at EziWith