Deep Image Harmonization in Dual Color Spaces


Image harmonization is an essential step in image composition that adjusts the appearance of composite foreground to address the inconsistency between foreground and background. Existing methods primarily operate in correlated 𝑅𝐺𝐡 color space, leading to entangled features and limited representation ability. In contrast, decorrelated color space (e.g., πΏπ‘Žπ‘) has decorrelated channels that provide disentangled color and illumination statistics. In this paper, we explore image harmonization in dual color spaces, which supplements entangled 𝑅𝐺𝐡 features with disentangled 𝐿, π‘Ž, 𝑏 features to alleviate the workload in harmonization process. The network comprises a 𝑅𝐺𝐡 harmonization backbone, an πΏπ‘Žπ‘ encoding module, and an πΏπ‘Žπ‘ control module. The backbone is a U-Net network translating composite image to harmonized image. Three encoders in πΏπ‘Žπ‘ encoding module extract three control codes independently from 𝐿, π‘Ž, 𝑏 channels, which are used to manipulate the decoder features in harmonization backbone via πΏπ‘Žπ‘ control module. Our code and model are available at

Proceedings of the 31th ACM International Conference on Multimedia (ACM MM 2023)