SCMIT: Style-Consistent Multi-Domain Image-to-Image Translation
Anh Q. Nguyen, Byeongkeun Kang, Yeejin Lee · IEEE Access · 2026
Multi-domain image-to-image translation has become increasingly prominent in computer vision, addressing the demand for versatile models that can transform images across various domains. However, current image-to-image translation methods are often limited to transforming images between specific domains, making it difficult to extend these models to additional domains. In addition, existing methods using disentangled representation learning often fail to accurately separate features, resulting in low-quality generated images. To address these challenges, we propose a novel framework for multi-domain image-to-image translation that uses multi-scale feature extraction to improve the quality of disentanglement learning. Our framework introduces a novel loss function, called style consistency loss, which more accurately maps style features to style space.We evaluate the effectiveness of our framework using standard evaluation metrics for image-to-image translation by comparing it to benchmark models. Experimental results show that our framework can produce high-quality, artifact-free images on various datasets.