Residual Fourier-Guided Module (RFGM)

• Residual Fourier-Guided Module is a frequency-domain restoration module that splits features into amplitude and phase for enhanced global illumination recovery.
• It employs residual guidance and channel-wise selection to preserve previous stage information and mitigate error accumulation in dark-image pipelines.
• RFGM is integrated as the first stage in dual-stage frameworks, enabling coarse brightness restoration before detailed refinement with lightweight computational overhead.

Searching arXiv for the cited papers and closely related terminology to ground the article in current sources. Residual Fourier-Guided Module (RFGM) denotes a frequency-domain restoration module in which Fourier decomposition is coupled with residual guidance to preserve informative priors across stages and channels. The term is explicitly introduced for extreme dark image restoration, where RFGM serves as the first-stage mechanism for global illumination recovery by separating feature representations into amplitude and phase, processing them independently, and recombining them through residual fusion (Zhang et al., 5 Aug 2025). In adjacent literature, closely related residual Fourier components appear under different names rather than as an explicit “RFGM”: FREDSR uses an FFC-based residual generator for single-image super-resolution, and Fourier residual networks use residual Fourier blocks for approximation theory (Woo et al., 2022); (Davis et al., 5 May 2026). This suggests that “RFGM” is both a specific module name in dark-image restoration and a broader design pattern in which Fourier-domain processing is stabilized by residual pathways.

1. Terminological scope and core definition

In "Beyond Illumination: Fine-Grained Detail Preservation in Extreme Dark Image Restoration" (Zhang et al., 5 Aug 2025), RFGM is defined as the first-stage module of a dual-stage framework for extremely dark images. Its stated role is to restore global illumination in the frequency domain while capturing inter-stage and inter-channel dependencies through residual connections. The module is motivated by a recurring failure mode in dark-image pipelines: sequential feature transformations can erase useful information, amplify unreliable priors, and accumulate errors across stages.

The same exact term does not appear in FREDSR or in the Fourier residual network theory paper. In FREDSR, however, the paper’s FFC-based residual generator embodies a related idea: local and global branches are processed separately, the global branch uses a real FFT2d, the two branches are fused, and the model predicts the residual between a bicubic-upsampled input and the ground truth (Woo et al., 2022). In "Fourier Residual Networks Achieve Spectral Accuracy for Discontinuous Functions" (Davis et al., 5 May 2026), the closest equivalent is the residual Fourier block

$$f_\ell(x)=f_{\ell-1}(x)+g_\ell(x)+h_\ell(f_{\ell-1}(x)),$$

where Fourier expansions are injected into the residual stream at every layer. A plausible implication is that the literature contains both a narrow, paper-specific use of RFGM and a broader residual-Fourier architectural family.

2. Frequency-domain workflow and mathematical formulation

In the dark-image restoration setting, RFGM operates on a feature map $F_{i-1} \in \mathbb{R}^{H \times W \times C}$ and applies FFT to obtain amplitude $A_{i-1}$ and phase $P_{i-1}$ (Zhang et al., 5 Aug 2025). Amplitude and phase are then processed separately using convolution and ReLU, after which they are recombined through residual guidance and returned to the spatial domain by iFFT, yielding $F_i$. The paper states that the first stage contains six identical RFGMs.

The amplitude branch is the explicit illumination-restoration path. The module flattens $A_{i-1}$ and $A_i$ to $\mathbb{R}^{HW \times C}$, computes a similarity matrix,

$M = MS(A_{i-1}, A_i),$

selects the most relevant channel index,

$V = \text{Top-1}(M),$

and generates a prior guidance map from the selected previous-stage amplitude channel,

$F_{i-1} \in \mathbb{R}^{H \times W \times C}$0

The guided residual amplitude update is then

$F_{i-1} \in \mathbb{R}^{H \times W \times C}$1

This is described in the source as channel-wise residual prior injection rather than direct replacement (Zhang et al., 5 Aug 2025).

The phase branch addresses structural compensation. The previous and current phase are concatenated and fused by convolution with a residual addition:

$F_{i-1} \in \mathbb{R}^{H \times W \times C}$2

The paper interprets amplitude mainly as brightness or illumination and phase as carrying much of the structural detail. RFGM therefore decomposes restoration into a global illumination path and a structural compensation path within the same Fourier-domain module (Zhang et al., 5 Aug 2025).

3. Residual guidance, dependency modeling, and error control

The defining property of RFGM is not merely Fourier processing but residual reuse of frequency-domain information. The paper states that useful information in dark-image frequency features is distributed across both stages and channels: earlier stages may encode stronger or cleaner illumination cues than later stages, and different channels may capture different contours or brightness distributions (Zhang et al., 5 Aug 2025). RFGM addresses these two forms of dependency explicitly.

For inter-stage dependency, the previous frequency representation is preserved and reused instead of being overwritten by the current transform. For inter-channel dependency, the Top-1 similarity selection identifies the most reliable amplitude prior channel and uses it to modulate the current amplitude estimate. The paper presents this as a way to avoid treating all channels equally when priors are weak or noisy.

The residual formulation is central to error mitigation. In the authors’ description, RFGM is designed to reduce error accumulation from pre-trained models and manual priors, which can be unreliable in very dark scenes (Zhang et al., 5 Aug 2025). Selective prior usage limits how much prior information is propagated, and residual fusion preserves the current estimate while allowing earlier-stage information to act as soft guidance. This suggests a conservative update mechanism: the module does not commit fully to the prior, but conditions the current frequency estimate on it.

An analogous residual logic appears in FREDSR, although not under the RFGM name. There, the network predicts the difference between the bicubic-upsampled input and the true image rather than reconstructing the full high-resolution image from scratch (Woo et al., 2022). The paper states: “we upscale the low-resolution input by using a bicubic upscaler. Then we use our model to predict the difference in value between the input bicubic upscaled image, and the true image.” In both cases, residual learning is used to prevent the model from relearning the identity mapping and to focus capacity on corrections.

4. Placement within larger architectures

RFGM is only the first stage of the dual-stage dark-image framework in which it was introduced (Zhang et al., 5 Aug 2025). Its output is intended to provide global illumination recovery and coarse structural integrity, after which a second stage performs detail refinement. The second stage contains two complementary components: Patch Mamba, which operates on channel-concatenated non-downsampled patches for pixel-level correlation modeling without resolution loss, and Grad Mamba, which focuses on high-gradient regions to prioritize sharp edges and boundaries. The paper assigns distinct responsibilities to the three modules: RFGM handles global brightness and structure stabilization, Patch Mamba performs local pixel-detail modeling, and Grad Mamba performs edge- and gradient-driven refinement.

The relationship to FREDSR is architectural rather than terminological. FREDSR is an encoder-decoder style image-to-image architecture adapted for single-image super-resolution, inspired by pix2pix and modified using Fast Fourier Convolutions (Woo et al., 2022). Its FFC block splits information into local and global channels, processes local information using traditional convolutions and global information using a real FFT2d, and combines both branches within each block. The generator bicubic-upsamples the low-resolution input, predicts a residual image through FFC-based layers, and adds that residual to the bicubic baseline. The discriminator is also trained on residuals: “We allow the discriminator to train to distinguish between the residuals, also known as the differences between the upscaled images and the true images, and our model's predicted residuals.” The paper does not name this design RFGM, but the correspondence is structurally direct.

In the theoretical Fourier residual network literature, the analogous module is more abstract (Davis et al., 5 May 2026). Each layer contains an upper branch driven by the raw input $F_{i-1} \in \mathbb{R}^{H \times W \times C}$3, a lower branch driven by the previous layer output $F_{i-1} \in \mathbb{R}^{H \times W \times C}$4, and a residual summation. The first layer is

$F_{i-1} \in \mathbb{R}^{H \times W \times C}$5

and subsequent layers are

$F_{i-1} \in \mathbb{R}^{H \times W \times C}$6

Here the Fourier-guided component is implemented through sine and cosine expansions rather than FFT/iFFT over image features. A plausible implication is that residual Fourier guidance can be instantiated either as a practical image-restoration block or as a constructive approximation-theoretic operator.

5. Empirical properties, efficiency, and modularity

The dark-image restoration paper emphasizes that RFGM is lightweight and can be seamlessly integrated into existing Fourier-based frameworks with minimal computational overhead (Zhang et al., 5 Aug 2025). Its stated ingredients are FFT/iFFT, simple convolutions, ReLU and sigmoid, similarity-based channel selection, and residual fusion. The plug-and-play claim is supported by experiments in which RFGM is inserted into FourLLIE and DMFourLLIE.

For FourLLIE, the reported baseline is 21.11 PSNR / 0.6256 SSIM / 0.1825 LPIPS with 0.12M parameters, while FourLLIE + RFGM achieves 21.39 / 0.6425 / 0.1683 with 0.126M parameters (Zhang et al., 5 Aug 2025). For DMFourLLIE, the baseline is 21.47 / 0.6331 / 0.1781 with 0.75M parameters, while DMFourLLIE + RFGM achieves 21.50 / 0.6408 / 0.1593 with 0.756M parameters. The ablation table also reports that removing RFGM lowers performance from the full model’s 21.59 / 0.6441 / 0.1607 to 21.24 / 0.6374 / 0.1696. The paper further states that six RFGMs are optimal in the component-count study.

FREDSR provides a related efficiency argument in the super-resolution domain (Woo et al., 2022). Its parameter count is 37,115, and the paper compares this with EDSR at 43 million parameters. On UHDSR4K 640×360 → 1920×1080 3× SR, FREDSR reports SSIM 0.883 and PSNR 27.776, compared with EDSR at SSIM 0.879 and PSNR 27.61, ESPCN at SSIM 0.858 and PSNR 27.02, and FSRCNN at SSIM 0.857 and PSNR 27.03. On UHDSR4K 1280×720 → 4K 3× SR, FREDSR reports SSIM 0.9645 and PSNR 35.3344. The paper attributes the efficiency of this Fourier-residual design to the FFC global receptive field, reduced need for large convolutional stacks, and residual learning.

The approximation-theoretic paper supplies a different form of empirical and theoretical support (Davis et al., 5 May 2026). For the sign function, the network is monotone, bounded, and overshoot-free, and the $F_{i-1} \in \mathbb{R}^{H \times W \times C}$7 error obeys

$F_{i-1} \in \mathbb{R}^{H \times W \times C}$8

For piecewise-$F_{i-1} \in \mathbb{R}^{H \times W \times C}$9 functions with one jump, the paper states

$A_{i-1}$0

These results do not concern image restoration, but they reinforce the broader residual-Fourier thesis: residual spectral refinement can be used to capture difficult structure efficiently.

6. Misconceptions, limitations, and broader significance

A common misconception is that RFGM simply denotes any use of FFT inside a neural network. The explicit RFGM formulation is narrower: it combines amplitude-phase decomposition, residual reuse across stages, selective channel-wise prior guidance, and recomposition by iFFT (Zhang et al., 5 Aug 2025). It is therefore not equivalent to a generic Fourier block, a pure FFT module, or an attention mechanism.

Another misconception is that the term can be retroactively applied without qualification to all residual Fourier models. FREDSR and Fourier residual networks do not use the name “Residual Fourier-Guided Module” in the papers themselves (Woo et al., 2022); (Davis et al., 5 May 2026). The connection is architectural and conceptual rather than terminological. In FREDSR, the closest match is the FFC-based residual generator; in the theory paper, it is the residual Fourier block built from sine and cosine activations. Precision on this point matters because the three works address different problems: extreme dark image restoration, single-image super-resolution, and approximation of discontinuous one-dimensional functions.

The principal limitation emphasized in the restoration literature is data and task specificity. FREDSR is described as achieving strong performance with “extreme parameter efficiency at the cost of per-dataset generalizeability,” and it performs worse on Manga109 and Urban100, which the paper attributes to distribution shift and the specialized nature of residuals (Woo et al., 2022). In the dark-image setting, the motivation for RFGM itself arises from unreliable priors and weak signals in extreme darkness, implying that prior selection and residual fusion are necessary precisely because the restoration regime is fragile (Zhang et al., 5 Aug 2025). In the approximation-theoretic setting, the scope is explicitly one-dimensional and constructive, focusing on functions on $A_{i-1}$1 and on expressive power rather than optimization (Davis et al., 5 May 2026).

Taken together, these papers position RFGM as a technically specific instance of a larger residual-Fourier paradigm. In the narrow sense, it is a first-stage frequency-domain illumination recovery module for extreme dark image restoration (Zhang et al., 5 Aug 2025). In the broader sense, it exemplifies a recurring architectural principle: Fourier-domain or Fourier-activated representations become more stable and more useful when embedded in residual pathways that preserve earlier information, separate global and local structure, and avoid replacing fragile intermediate estimates outright (Woo et al., 2022); (Davis et al., 5 May 2026).