CycleGAN for day-to-night image translation: a comparative study

Raihan Taufiq, Muhammad Feriansyah; Rahadianti, Laksmita

doi:10.11591/ijai.v14.i3.pp2347-2357

Open access Journal SINTA 1

CycleGAN for day-to-night image translation: a comparative study

IAES International Journal of Artificial Intelligence · Vol. 14 No. 3 (2025) · pp. 2347–2357

DOI: 10.11591/ijai.v14.i3.pp2347-2357
ISSN: 2252-8938, 2089-4872
Published: 2025-06-01
Publisher: Institute of Advanced Engineering and Science
Language: eng
License: creativecommons.org/licenses/by-sa/4.0

Abstract

Computer vision tasks often fail when applied to night images, because the models are usually trained using clear daytime images only. This creates the need to augment the data with more nighttime image for training to increase robustness. In this study, we consider day-to-night image translation using both traditional image processing approaches and deep learning models. This study employs a hybrid framework of traditional image processing followed by a CycleGANbased deep learning model for day-to-night image translation. We then conduct a comparative study on various generator architectures in our CycleGAN model. This research compares four different CycleGAN models; i.e., the orginal CycleGAN, feature pyramid network (FPN) based CycleGAN, the original U-Net vision transformer based UVCGAN, plus a modified UVCGAN with additional edge loss. The experimental results show that the orginal UVCGAN obtains an Frechet inception distance (FID) score of 16.68 and structural similarity index ´ measure (SSIM) of 0.42, leading in terms of FID. Meanwhile, FPN-CycleGAN obtains an FID score of 104.46 and SSIM score of 0.44, leading in terms of SSIM. Considering FPN-CycleGAN’s bad FID score and visual observation, we conclude that UVCGAN is more effective in generating synthetic nighttime images.

Keywords

Collaboration network of the authors

8 authors · 10 co-authorship links

Hover or tap a name to highlight its co-authors. Double-click to open the author's page.

University of Indonesia
Affiliation unknown

Different organisations
Same organisation
Affiliation unknown
Thicker line = more joint articles · bigger circle = more articles

Show as table

Co-author	Organisation	Joint articles
Made Prastha Nugraha	University of Indonesia	1
Ikhlasul Akmal Hanif	University of Indonesia	1
Eduardus Tjitrahardja	University of Indonesia	1
Rahmat Bryan Naufal	University of Indonesia	1
Steven Christ Pinantyo Arwidarasto	University of Indonesia	1
Alexander Calvin	University of Indonesia	1

Affiliations come from OpenAlex; journal OAI-PMH metadata carries none. Authors without an OpenAlex ID are matched by name, which can merge or split people.

Metadata source: OAI-PMH (metadata jurnal) http://ijai.iaescore.com/index.php/IJAI/oai, first harvested 8 Nov 2026, last seen 8 Nov 2026.

Original metadata (oai_dc)

<oai_dc:dc xmlns="http://www.openarchives.org/OAI/2.0/" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:oai_dc="http://www.openarchives.org/OAI/2.0/oai_dc/" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://www.openarchives.org/OAI/2.0/oai_dc/  http://www.openarchives.org/OAI/2.0/oai_dc.xsd">
	<dc:title xml:lang="en-US">CycleGAN for day-to-night image translation: a comparative study</dc:title>
	<dc:creator>Raihan Taufiq, Muhammad Feriansyah</dc:creator>
	<dc:creator>Rahadianti, Laksmita</dc:creator>
	<dc:subject xml:lang="en-US"></dc:subject>
	<dc:subject xml:lang="en-US">CycleGAN; Deep learning; Image processing; Image translation; Synthetic nighttime images</dc:subject>
	<dc:description xml:lang="en-US">Computer vision tasks often fail when applied to night images, because the models are usually trained using clear daytime images only. This creates the need to augment the data with more nighttime image for training to increase robustness. In this study, we consider day-to-night image translation using both traditional image processing approaches and deep learning models. This study employs a hybrid framework of traditional image processing followed by a CycleGANbased deep learning model for day-to-night image translation. We then conduct a comparative study on various generator architectures in our CycleGAN model. This research compares four different CycleGAN models; i.e., the orginal CycleGAN, feature pyramid network (FPN) based CycleGAN, the original U-Net vision transformer based UVCGAN, plus a modified UVCGAN with additional edge loss. The experimental results show that the orginal UVCGAN obtains an Frechet inception distance (FID) score of 16.68 and structural similarity index ´ measure (SSIM) of 0.42, leading in terms of FID. Meanwhile, FPN-CycleGAN obtains an FID score of 104.46 and SSIM score of 0.44, leading in terms of SSIM. Considering FPN-CycleGAN’s bad FID score and visual observation, we conclude that UVCGAN is more effective in generating synthetic nighttime images.</dc:description>
	<dc:publisher xml:lang="en-US">Institute of Advanced Engineering and Science</dc:publisher>
	<dc:contributor xml:lang="en-US">Faculty of Computer Science, Universitas Indonesia</dc:contributor>
	<dc:date>2025-06-01</dc:date>
	<dc:type>info:eu-repo/semantics/article</dc:type>
	<dc:type>info:eu-repo/semantics/publishedVersion</dc:type>
	<dc:type xml:lang="en-US"></dc:type>
	<dc:format>application/pdf</dc:format>
	<dc:identifier>https://ijai.iaescore.com/index.php/IJAI/article/view/27081</dc:identifier>
	<dc:identifier>10.11591/ijai.v14.i3.pp2347-2357</dc:identifier>
	<dc:source xml:lang="en-US">IAES International Journal of Artificial Intelligence (IJ-AI); Vol 14, No 3: June 2025; 2347-2357</dc:source>
	<dc:source>2252-8938</dc:source>
	<dc:source>2089-4872</dc:source>
	<dc:source>10.11591/ijai.v14.i3</dc:source>
	<dc:language>eng</dc:language>
	<dc:relation>https://ijai.iaescore.com/index.php/IJAI/article/view/27081/14575</dc:relation>
	<dc:rights xml:lang="en-US">Copyright (c) 2025 Institute of Advanced Engineering and Science</dc:rights>
	<dc:rights xml:lang="en-US">http://creativecommons.org/licenses/by-sa/4.0</dc:rights>
</oai_dc:dc>