Attention-Enhanced CycleGAN for Unpaired Image-to-Image Translation

DOI

10.2991/978-94-6239-648-7_100How to use a DOI?

Keywords

Unpaired translation; CycleGAN; Frozen attention; Mask-guided discriminator; Foreground-aware translation

Abstract

Cycle-consistent GANs are widely used for unpaired image-to-image translation, but they often over-translate textures in regions that should remain largely unchanged (e.g., background grass or sky in horse ↔ zebra). This failure mode is encouraged by discriminators that score the full image uniformly, which can reward the generator for spreading domain-specific cues beyond the intended foreground. The paper proposes a lightweight, plug-in attention head attached to the CycleGAN generator to predict a soft foreground attention mask. During training the paper uses this mask to (i) gate the discriminator’s input and (ii) weight the adversarial loss so that discriminator feedback is concentrated on attended regions. A key stability issue in attention-guided adversarial training is that attention maps can drift late in training; to prevent this, the paper monitor attention-map changes across epochs and freeze the attention parameters once the maps stabilize. On the Horse ↔ Zebra benchmark, our method improves Kernel Inception Distance (KID; lower is better) compared with a vanilla CycleGAN baseline and yields visibly cleaner backgrounds with reduced texture bleeding.

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  - Wenqi Zheng
PY  - 2026
DA  - 2026/04/24
TI  - Attention-Enhanced CycleGAN for Unpaired Image-to-Image Translation
BT  - Proceedings of the International Workshop on Advances in Deep Learning for Image Analysis and Computer Vision (IWADIC 2025)
PB  - Atlantis Press
SP  - 934
EP  - 942
SN  - 2352-538X
UR  - https://doi.org/10.2991/978-94-6239-648-7_100
DO  - 10.2991/978-94-6239-648-7_100
ID  - Zheng2026
ER  -