A Novel Image-to-Image Model: MSF-CycleGAN

DOI

10.2991/978-94-6239-648-7_97How to use a DOI?

Keywords

Multi-scale feature consistency; Unpaired image-to-image translation; CycleGAN; Feature-level cycle constraint; VGG perceptual features

Abstract

Unpaired image-to-image transformations are often affected by structural distortions and semantic inconsistencies because they rely on pixel-level cyclic consistency constraints. To address these limitations, the paper proposes a multi-scale feature consistency cyclic generative adversarial network, which introduces cyclic constraints in the feature space of the pre-trained Visual Geometry Group (VGG) network. By imposing constraints simultaneously at the low-level structural edges, intermediate textures, and high-level semantic representations, the proposed method enhances the retention of structure while reducing artifacts and background shifts during the transformation process. Additionally, an adaptive weighting mechanism is introduced to automatically balance the influence of different feature scales, enabling the model to capture finer textures while maintaining the integrity of the overall semantic layout. Experimental results on the horse2zebra dataset show that compared to the standard Cycle-Consistent Generative Adversarial Network (CycleGAN), the proposed framework generates clearer stripe textures, more coherent body contours, and better semantic consistency. These advancements demonstrate the robustness of multi-scale feature guidance to stabilize unpaired translation. Going forward, this method offers a promising basis to combine domain-adaptive perceptual features with more computationally efficient feature-level constraints and to evolve feature-level constraints for applications of unpaired translation into a wider application in complex real-world scenarios.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Yujing Wang
PY  - 2026
DA  - 2026/04/24
TI  - A Novel Image-to-Image Model: MSF-CycleGAN
BT  - Proceedings of the International Workshop on Advances in Deep Learning for Image Analysis and Computer Vision (IWADIC 2025)
PB  - Atlantis Press
SP  - 903
EP  - 912
SN  - 2352-538X
UR  - https://doi.org/10.2991/978-94-6239-648-7_97
DO  - 10.2991/978-94-6239-648-7_97
ID  - Wang2026
ER  -