Fd-CycleGAN: Frequency Distributed CycleGAN

• The paper introduces Fd-CycleGAN, which integrates Local Neighborhood Encoding and frequency-aware supervision to enhance latent representation learning and mitigate mode collapse.
• It utilizes divergence-based losses like KL/JS divergence and log-based similarity measures to align spatial and frequency distributions, resulting in faster convergence and improved perceptual quality.
• Fd-CycleGAN is applied to tasks such as document restoration, artistic style transfer, and medical image synthesis, showing superior performance particularly in low-data regimes.

Fd-CycleGAN, expanded as Frequency Distributed CycleGAN, is an image-to-image translation framework that extends the CycleGAN paradigm by enhancing latent representation learning to approximate real data distributions. Its defining additions are Local Neighborhood Encoding (LNE) and frequency-aware supervision, introduced to capture fine-grained local pixel semantics while preserving structural coherence from the source domain. The framework employs distribution-based loss metrics, including KL/JS divergence and log-based similarity measures, to quantify alignment between real and generated image distributions in both spatial and frequency domains. Reported evaluations on Horse2Zebra, Monet2Photo, and a synthetically augmented Strike-off dataset indicate superior perceptual quality, faster convergence, and improved mode diversity relative to baseline CycleGAN and other state-of-the-art methods, particularly in low-data regimes (Nigam et al., 5 Aug 2025).

1. Conceptual basis and problem setting

Fd-CycleGAN is motivated by limitations attributed to regular CycleGAN, specifically semantic ambiguity, loss of local structure, and mode collapse, which are described as particularly acute in data-scarce or structurally complex scenarios such as document artifacts, fine artistic styles, or medical imaging (Nigam et al., 5 Aug 2025).

The framework retains the standard bi-directional CycleGAN backbone with two generators, $\mathcal{G}_{X \rightarrow Y}$ and $\mathcal{G}_{Y \rightarrow X}$, and two discriminators, $\mathcal{D}_X$ and $\mathcal{D}_Y$. Its novelty lies not in replacing this adversarial translation scaffold, but in augmenting it with mechanisms intended to improve how latent and reconstructed representations reflect both local semantics and distributional structure. In the terminology of the source work, the method seeks to capture both local and global distribution characteristics, thereby producing translations that are more visually coherent and semantically consistent.

This design positions Fd-CycleGAN as a CycleGAN derivative concerned with representation quality and distribution alignment rather than with eliminating cycle consistency, changing the adversarial game, or replacing the backbone generators outright. That positioning is explicit in the claim that LNE preprocessing and frequency-aware supervision are plug-and-play additions that do not alter the intrinsic generator or discriminator architectures.

2. Architectural composition and Local Neighborhood Encoding

The architectural core of Fd-CycleGAN remains the standard CycleGAN formulation, but the input stream is enriched by Local Neighborhood Encoding. LNE is described as a preprocessing module that encodes each pixel according to spectral similarity with its two-hop neighbors using Gaussian weighting (Nigam et al., 5 Aug 2025).

For a pixel $x_i$ and neighbors $x_j$, the neighborhood weights are given as

$$w_{j|i} = \exp\left(-\frac{\|x_i - x_j\|^2}{2\sigma_i^2}\right)$$

and the normalized neighborhood relation as

$p_{j|i} = \frac{w_{j|i}}{\sum_{j \neq i} w_{j|i}}.$

In the supplied description, this representation smoothes noise, reduces abrupt intensity changes, and accentuates semantically meaningful local variations. The intended effect is to enrich the translated signal with local spatial context before adversarial and cycle-based constraints act on it. Because LNE is a preprocessing stage, it modifies images prior to their passage through the generator and discriminator rather than changing the internal topology of those networks.

The paper further characterizes the generator side as standard CycleGAN (ResNet-type), emphasizing ease of transition from vanilla CycleGAN. The resulting system therefore combines conventional unpaired translation machinery with an explicit local-context encoder. This suggests a deliberate attempt to improve translation fidelity without incurring the implementation complexity associated with a wholly new generator family.

3. Frequency-aware supervision and distributional objectives

The second major component of Fd-CycleGAN is frequency-aware supervision, which supervises generated outputs through multiple statistical summaries rather than through pixel correspondence alone (Nigam et al., 5 Aug 2025). The framework uses five distribution functions:

Distribution function	Brief role	Granularity
Gaussian distribution	Local-window estimate	Local
Histogram distribution	Intensity counts in window	Local
Weighted histogram	Center-weighted histogram	Local
Categorical distribution	Normalized intensity frequency	Global/local
Patch-wise categorical	Categorical over non-overlapping patches	Patch

These functions provide “different views of local/global statistics,” with varying levels of granularity determined by kernel or patch size. The conceptual move is from raw reconstruction toward distribution alignment: generated and target images are compared not only as arrays of pixels but also as carriers of local and global statistical structure.

The source contrasts this with the standard CycleGAN cycle loss, written there as an $L_1$ reconstruction term, and states that Fd-CycleGAN replaces the cycle-consistency loss with a divergence-based form:

$$\mathcal{L}_{cyc}^{div} = StD(\mathcal{X}_{real} \Vert \mathcal{X}_{rec}) + StD(\mathcal{Y}_{real} \Vert \mathcal{Y}_{rec}).$$

The framework employs KL divergence, JS divergence, and a log-based loss for distributional comparison. The log-based term is given as

$\mathcal{G}_{Y \rightarrow X}$0

The reported total objective is

$\mathcal{G}_{Y \rightarrow X}$1

Identity and adversarial losses are stated to remain the same as in original CycleGAN. The resulting optimization scheme is intended to enforce closeness in both spatial and frequency domains, thereby addressing the source paper’s critique that pure pixel-wise cycle consistency is “only pixel-wise, not distribution-aligning.”

4. Experimental characterization

The experimental program reported for Fd-CycleGAN spans Horse2Zebra, Monet2Photo, and a synthetically augmented Strike-off dataset. The paper states that the method demonstrates superior perceptual quality, faster convergence, and improved mode diversity relative to baseline CycleGAN and other state-of-the-art methods, with particular strength in low-data regimes (Nigam et al., 5 Aug 2025).

A quantitative comparison provided for Horse→Zebra is summarized below.

Model variant	PSNR / SSIM	FID / CLIP
CycleGAN (baseline)	18.53 / 0.67	77.18 / 28.07
Fd-CycleGAN (L1 loss, wt. image)	25.98 / 0.89	81.67 / 29.13
Fd-CycleGAN (Hist+Catg, JSD, wt. im)	24.54 / 0.86	67.53 / 22.11
Fd-CycleGAN (Gauss+Catg, L1)	25.91 / 0.88	72.82 / 28.91
CUT (Contrastive)	13.71 / 0.35	45.50 / 29.15
Diffusion (CycleDiff)	11.51 / 0.21	347.27 / 25.04
CycleNet (Diffusion)	20.42 / 0.52	81.69 / 28.91

Several patterns are explicitly identified in the source. First, Fd-CycleGAN is described as substantially improving PSNR and SSIM, especially with L1 or JSD loss and categorical or frequency features. Second, the JSD variant is reported to yield the best FID among the Fd-CycleGAN variants listed. Third, the paper states that weighted image variants always outperform no-weighting versions, with a statistical test showing $\mathcal{G}_{Y \rightarrow X}$2 for improvement in FID and SSIM. Fourth, the log-based loss is reported to give the quickest convergence, whereas divergence-based losses incur mildly higher computational cost but provide superior sample diversity and stability, specifically in relation to mode collapse.

The paper also states that Fd-CycleGAN models are more efficient than diffusion models, “as much as an order of magnitude in runtime,” while domain complexity affects convergence time; Monet2Photo is noted as requiring longer convergence than strike-off removal.

5. Position within the CycleGAN variant literature

Fd-CycleGAN belongs to a broader line of CycleGAN adaptations in which specific deficiencies of unpaired translation are addressed through additional constraints, modules, or architectural reorganizations. Its distinctive intervention is frequency-guided latent learning combined with local-context encoding, rather than generator redesign alone or task-specific structural priors.

Other CycleGAN variants in the supplied literature modify the framework in different directions. Feature Cycling Block introduces a dual-branch generator unit with memory and image branches for GAN image synthesis, improving FID on datasets including FFHQ and LSUN without changing objective functions or training skills (Park et al., 2022). Vessel-Segmentation-Based CycleGAN inserts a pretrained vessel segmentation network before each generator and adds a Dice-based segmentation loss to preserve retinal vessel locations in unpaired multimodal retinal synthesis (Sindel et al., 2023). Identity-aware CycleGAN extends CycleGAN to image-to-video and video-to-image face translation by combining WGAN-GP with a pretrained FaceNet-based identity loss (Huang et al., 2017). Federated CycleGAN decomposes CycleGAN loss into domain-specific local objectives to enable privacy-preserving federated training, and further introduces a switchable AdaIN-based architecture to reduce communication cost (Song et al., 2021). Cycle-free CycleGAN replaces the two-generator/two-discriminator setup with a single invertible generator and one discriminator, using invertibility to guarantee cycle consistency in low-dose CT denoising (Kwon et al., 2021).

Against this backdrop, Fd-CycleGAN is distinguished by the claim that its enhancements are plug-and-play and do not alter intrinsic generator or discriminator architectures. This suggests that it is best understood as a distributional supervision layer over vanilla CycleGAN, rather than as an architectural overhaul comparable to invertible, federated, segmentation-conditioned, or identity-aware variants.

6. Applications, interpretation, and nomenclature

The applications identified for Fd-CycleGAN are document restoration, artistic style transfer, and medical image synthesis (Nigam et al., 5 Aug 2025). In document restoration, the method is described as excelling at strike-off/ink noise removal by leveraging local structure and statistical alignment. In artistic style transfer, especially on Monet2Photo and related style datasets, the framework is presented as enforcing both global semantics and local detail, including texture and brushstroke statistics. In medical imaging, the source emphasizes suitability for limited data regimes, where generalization and structure preservation are critical.

The abstract states that the reported results “suggest that frequency-guided latent learning significantly improves generalization in image translation tasks,” and that the framework provides comparative insights against diffusion-based generative models, emphasizing the advantages of a lightweight adversarial approach in training efficiency and qualitative output. Because the paper ties these claims to low-data settings, a plausible implication is that Fd-CycleGAN is intended not only as a quality improvement over baseline CycleGAN but also as a pragmatic alternative where sample scarcity makes purely data-hungry generative strategies less attractive.

A point of nomenclature merits clarification. In the supplied literature, the explicit title “Learning Latent Representations for Image Translation using Frequency Distributed CycleGAN” uses Fd-CycleGAN to mean Frequency Distributed CycleGAN (Nigam et al., 5 Aug 2025). Separately, a dehazing, de-snowing, and de-raining paper describes a hybrid FFA + CycleGAN system as analogous to “Fd-CycleGAN” while focusing on attention-based feature fusion and paired/unpaired restoration (Jain, 8 Mar 2025). This suggests that the abbreviation Fd is not fully self-disambiguating across adjacent works and benefits from explicit expansion in technical citation contexts.