WaveMamba Fusion Block (WMFB)

• WMFB is a module that decomposes RGB and infrared features into low- and high-frequency sub-bands using Discrete Wavelet Transform for effective cross-modal fusion.
• It fuses low-frequency components through the LMFB, employing channel swapping, VSS blocks, and gated attention, while high-frequency data is merged with an absolute-maximum rule.
• The integrated IDWT-based detection head reconstructs features efficiently, contributing to performance gains (up to 4.5% mAP improvement) and reduced parameters.

WaveMamba Fusion Block (WMFB) is the core fusion module introduced in “WaveMamba: Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection” for cross-modality fusion in the wavelet domain. In that formulation, WMFB separates RGB and infrared features into low- and high-frequency sub-bands by Discrete Wavelet Transform (DWT), fuses the low-frequency component with a Mamba-based mechanism, fuses the high-frequency components by an absolute-maximum rule, and supplies the resulting sub-bands to an Inverse Discrete Wavelet Transform (IDWT)-based detection head (Zhu et al., 24 Jul 2025). In a broader research context, the label is not standardized: several related papers use different module names for closely related wavelet-guided state-space fusion mechanisms, and some do not use the term “WMFB” at all (Cai et al., 19 Dec 2025, Wang et al., 15 Aug 2025, Zhang et al., 5 Aug 2025, Ahmad et al., 2024, Liu et al., 3 Mar 2026).

1. Terminology and scope

The strict, paper-specific meaning of WMFB comes from WaveMamba for RGB-infrared object detection, where it is explicitly named and described as the module that “facilitates comprehensive fusion across low-/high-frequency sub-bands” (Zhu et al., 24 Jul 2025). Outside that paper, the term is used much less consistently. Several later or related works describe functionally similar wavelet-integrated Mamba or state-space fusion units under different names, while one infrared-visible fusion paper provides insufficient extracted technical text to confirm whether “WMFB” appears at all (Zhang et al., 24 Mar 2025).

Source	Exact term in source	Relation to WMFB
(Zhu et al., 24 Jul 2025)	WMFB	Explicitly introduced
(Zhang et al., 24 Mar 2025)	Wavelet-SSM module; cross-modal feature attention modulation	WMFB-specific details cannot be confirmed
(Cai et al., 19 Dec 2025)	WHF; DAFF	Closest functional match, but no WMFB term
(Wang et al., 15 Aug 2025)	HFEB; mask-guided fusion	Treated as WMFB for summary purposes, not explicit
(Zhang et al., 5 Aug 2025)	MDIF	Mapped directly to WMFB in the technical description
(Ahmad et al., 2024)	Wavelet Transformation stage	Fusion stage mapped to WMFB
(Liu et al., 3 Mar 2026)	WSSB; Freq-SSM	Closest practical equivalent

A common misconception is that WMFB denotes a universally standardized module family. The available papers do not support that reading. The exact name is explicit in the object-detection model of WaveMamba, whereas related segmentation, restoration, and hyperspectral-classification papers describe analogous constructs with task-specific nomenclature and architectural variations (Zhu et al., 24 Jul 2025, Cai et al., 19 Dec 2025, Zhang et al., 5 Aug 2025).

2. Canonical formulation in WaveMamba

In its canonical form, WMFB is embedded in a dual-stream RGB-infrared detector. Features from early backbone stages are decomposed via DWT into four sub-bands per modality: one low-frequency sub-band, $LL$, and three high-frequency sub-bands, $LH$, $HL$, and $HH$ (Zhu et al., 24 Jul 2025). The design is driven by a specific empirical observation reported for the M3FD validation set: normalized information entropy is higher in infrared for low-frequency bands and higher in RGB for high-frequency bands. The paper interprets this as indicating that infrared preserves more semantics and shapes at coarse scales, while RGB encodes sharper local edges and textures (Zhu et al., 24 Jul 2025).

Using Haar wavelets, the default 2D DWT is defined with

$$L = \frac{1}{\sqrt{2}}[1,1], \qquad H = \frac{1}{\sqrt{2}}[1,-1].$$

For a feature map $I$, the four sub-bands are

$$LL = (I \star l_x \star l_y)\downarrow 2,\quad LH = (I \star l_x \star h_y)\downarrow 2,\quad HL = (I \star h_x \star l_y)\downarrow 2,\quad HH = (I \star h_x \star h_y)\downarrow 2,$$

where $\star$ denotes convolution and $\downarrow 2$ denotes downsampling by 2 in both spatial dimensions (Zhu et al., 24 Jul 2025). For an input tensor $f^i$ of shape $LH$0, each sub-band has shape $LH$1 (Zhu et al., 24 Jul 2025).

WMFB then decouples fusion into two regimes. Low-frequency sub-bands are fused by the Low-frequency Mamba Fusion Block (LMFB), whereas high-frequency sub-bands are fused independently by a parameter-free high-frequency enhancement rule (Zhu et al., 24 Jul 2025). The fused sub-bands are subsequently reconstructed in an improved detection head that replaces conventional upsampling with IDWT, thereby reducing information loss (Zhu et al., 24 Jul 2025).

Architecturally, WaveMamba applies DWT on the second-stage features $LH$2 and $LH$3, then uses multi-level DWT and WMFB on fused features at the 3rd and 5th layers. Ablation identifies three WMFBs placed at stages $LH$4 as the best configuration (Zhu et al., 24 Jul 2025).

3. Low-frequency fusion through LMFB

The low-frequency branch of WMFB is the Low-frequency Mamba Fusion Block, or LMFB. It is built from two stages: Shallow Fusion Mamba (SFM) and Deep Fusion Mamba (DFM) (Zhu et al., 24 Jul 2025). SFM performs channel swapping followed by Visual State Space (VSS) blocks, while DFM performs selective state-space fusion with gated attention to suppress redundancy and enhance complementary content (Zhu et al., 24 Jul 2025).

In channel swapping, the low-frequency RGB and infrared features are split along the channel dimension and partially exchanged. A simple instantiation given in the paper is

$LH$5

then

$LH$6

The resulting tensors are processed by VSS blocks,

$LH$7

which apply SS2D scanning with selective state updates (Zhu et al., 24 Jul 2025).

The DFM stage uses two-stream processing per modality. For the primary modality $LH$8 and auxiliary modality $LH$9, the paper describes a primary stream with $HL$0 depth-wise convolution, SiLU, SS2D, and LayerNorm producing $HL$1, an auxiliary stream producing $HL$2, and a gate $HL$3 that regulates both streams. A representative formulation consistent with the paper is

$HL$4

optionally with a residual connection (Zhu et al., 24 Jul 2025). This is the point at which WMFB most clearly expresses its Mamba-based character: SS2D scans images in four directions to construct sequences tailored to 2D structure, and LayerNorm stabilizes training (Zhu et al., 24 Jul 2025).

The underlying state-space principle is the generic discrete SSM

$HL$5

which, in VMamba’s SS2D form, is applied to image-derived sequences rather than to a single rasterized token order (Zhu et al., 24 Jul 2025). The intended effect is efficient linear-time modeling of long-range low-frequency semantics without the quadratic cost of self-attention (Zhu et al., 24 Jul 2025).

4. High-frequency fusion and inverse-wavelet reconstruction

WMFB treats high-frequency information differently from low-frequency semantics. Rather than applying another learned Mamba-style fusion stage, WaveMamba fuses each high-frequency sub-band independently by an absolute-maximum rule, denoted HFE in the technical description (Zhu et al., 24 Jul 2025). For a high-frequency sub-band $HL$6, the fused output is

$HL$7

Equivalently, the implementation is described by binary masks:

$HL$8

where $HL$9 produces 1 where the argument is positive and 0 otherwise (Zhu et al., 24 Jul 2025).

This choice reflects a second architectural distinction. A common assumption is that wavelet-Mamba fusion blocks must learn both low- and high-frequency combination. In the original WMFB, that is not the case. Low-frequency fusion is learned and state-space driven; high-frequency fusion is parameter-free and explicitly preserves the strongest edge or detail response at each location (Zhu et al., 24 Jul 2025).

After fusion, the improved YOLOv8 head reconstructs higher-resolution features by IDWT rather than by nearest or bilinear interpolation. A canonical synthesis formula given in the paper is

$HH$0

where $HH$1 denotes upsampling by 2 (Zhu et al., 24 Jul 2025). The rest of the YOLOv8 head remains unchanged; only the upsampling paths are upgraded to IDWT (Zhu et al., 24 Jul 2025). This makes WMFB not merely a fusion unit but part of a larger wavelet-consistent analysis–synthesis pipeline.

5. Empirical profile, ablations, and efficiency

WaveMamba reports that it surpasses state-of-the-art approaches with an average mAP improvement of about 4.5% across four benchmarks (Zhu et al., 24 Jul 2025). Representative results with the YOLOv8 backbone are reported as follows: on M3FD, $HH$2 and $HH$3; on DroneVehicle, $HH$4 and $HH$5; on LLVIP, $HH$6 and $HH$7; and on FLIR-Aligned, $HH$8 and $HH$9 (Zhu et al., 24 Jul 2025).

The ablation studies isolate the contribution of each WMFB component. On M3FD with the YOLOv8 backbone, removing SFM produces a drop of $$L = \frac{1}{\sqrt{2}}[1,1], \qquad H = \frac{1}{\sqrt{2}}[1,-1].$$0 in mAP50 and $$L = \frac{1}{\sqrt{2}}[1,1], \qquad H = \frac{1}{\sqrt{2}}[1,-1].$$1 in mAP, while removing DFM produces a drop of $$L = \frac{1}{\sqrt{2}}[1,1], \qquad H = \frac{1}{\sqrt{2}}[1,-1].$$2 in mAP50 and $$L = \frac{1}{\sqrt{2}}[1,1], \qquad H = \frac{1}{\sqrt{2}}[1,-1].$$3 in mAP (Zhu et al., 24 Jul 2025). A fair comparison of fusion strategies with the original head yields the following sequence of results: baseline without DWT, $$L = \frac{1}{\sqrt{2}}[1,1], \qquad H = \frac{1}{\sqrt{2}}[1,-1].$$4; $$L = \frac{1}{\sqrt{2}}[1,1], \qquad H = \frac{1}{\sqrt{2}}[1,-1].$$5, $$L = \frac{1}{\sqrt{2}}[1,1], \qquad H = \frac{1}{\sqrt{2}}[1,-1].$$6; $$L = \frac{1}{\sqrt{2}}[1,1], \qquad H = \frac{1}{\sqrt{2}}[1,-1].$$7, $$L = \frac{1}{\sqrt{2}}[1,1], \qquad H = \frac{1}{\sqrt{2}}[1,-1].$$8; $$L = \frac{1}{\sqrt{2}}[1,1], \qquad H = \frac{1}{\sqrt{2}}[1,-1].$$9, $I$0; and $I$1, $I$2. Adding the improved head gives the final $I$3 result of $I$4 (Zhu et al., 24 Jul 2025). The improved head alone contributes $I$5 mAP50 and $I$6 mAP while reducing parameters by $I$7M (Zhu et al., 24 Jul 2025).

Wavelet-basis sensitivity is reported to be limited. Performance is robust across MEYR, SYM3, COIF3, and DB3, with Haar achieving the best result of $I$8; the differences are reported as $I$9 mAP50 and $$LL = (I \star l_x \star l_y)\downarrow 2,\quad LH = (I \star l_x \star h_y)\downarrow 2,\quad HL = (I \star h_x \star l_y)\downarrow 2,\quad HH = (I \star h_x \star h_y)\downarrow 2,$$0 mAP (Zhu et al., 24 Jul 2025). With the YOLOv8 backbone, the full model has parameters of approximately $$LL = (I \star l_x \star l_y)\downarrow 2,\quad LH = (I \star l_x \star h_y)\downarrow 2,\quad HL = (I \star h_x \star l_y)\downarrow 2,\quad HH = (I \star h_x \star h_y)\downarrow 2,$$1M and mean inference time of approximately $$LL = (I \star l_x \star l_y)\downarrow 2,\quad LH = (I \star l_x \star h_y)\downarrow 2,\quad HL = (I \star h_x \star l_y)\downarrow 2,\quad HH = (I \star h_x \star h_y)\downarrow 2,$$2 ms for $$LL = (I \star l_x \star l_y)\downarrow 2,\quad LH = (I \star l_x \star h_y)\downarrow 2,\quad HL = (I \star h_x \star l_y)\downarrow 2,\quad HH = (I \star h_x \star h_y)\downarrow 2,$$3 pairs on an A800 GPU (Zhu et al., 24 Jul 2025). On FLIR-Aligned, the paper states that WaveMamba yields higher accuracy with fewer parameters and faster inference than attention-heavy fusion baselines such as CrossFormer (Zhu et al., 24 Jul 2025).

These results position WMFB as a module whose empirical contribution depends on the coordinated action of three elements: LMFB for low-frequency semantics, HFE for high-frequency detail preservation, and IDWT-based reconstruction in the head (Zhu et al., 24 Jul 2025).

6. Related formulations, extensions, and naming ambiguity

Several subsequent works adopt closely related design principles while avoiding the exact term WMFB. In “WDFFU-Mamba,” the paper states that it does not use the term “WaveMamba Fusion Block (WMFB)” and instead introduces a Wavelet-denoised High-Frequency-guided Feature (WHF) module and a Dual Attention Feature Fusion (DAFF) module. Functionally, DAFF integrating WHF at the bottleneck is described as the closest match to what one might call a WMFB (Cai et al., 19 Dec 2025). In “Guiding WaveMamba with Frequency Maps for Image Debanding,” the paper likewise does not use the term explicitly; for summary purposes, the High-Frequency Enhance Block (HFEB), together with mask-guided fusion into inverse wavelet reconstruction, is treated as the WMFB-like component (Wang et al., 15 Aug 2025).

A similar pattern appears in three-dimensional and hyperspectral settings. “SSFMamba” states that the paper does not explicitly define a WMFB and instead names its fusion unit the Multi-domain Information Fusion Block (MDIF), with a 3D Multi-Directional Scanning Mechanism (MDSM) inside a Mamba block (Zhang et al., 5 Aug 2025). “WaveMamba: Spatial-Spectral Wavelet Mamba for Hyperspectral Image Classification” also states that the paper does not explicitly define a module named WMFB; the relevant fusion is implemented by a wavelet transformation stage together with a spatial and spectral gate and concatenation of sub-bands into a fused tensor $$LL = (I \star l_x \star l_y)\downarrow 2,\quad LH = (I \star l_x \star h_y)\downarrow 2,\quad HL = (I \star h_x \star l_y)\downarrow 2,\quad HH = (I \star h_x \star h_y)\downarrow 2,$$4 that is passed to the Mamba/state-space model (Ahmad et al., 2024). “CAWM-Mamba” again does not name a module WMFB; the closest fully specified components are the Cross-modal Feature Interaction Module (CFIM), the Wavelet Space State Block (WSSB), and its Frequency-Selective State Space Module (Freq-SSM), with a practical WMFB described as CFIM $$LL = (I \star l_x \star l_y)\downarrow 2,\quad LH = (I \star l_x \star h_y)\downarrow 2,\quad HL = (I \star h_x \star l_y)\downarrow 2,\quad HH = (I \star h_x \star h_y)\downarrow 2,$$5 WSSB $$LL = (I \star l_x \star l_y)\downarrow 2,\quad LH = (I \star l_x \star h_y)\downarrow 2,\quad HL = (I \star h_x \star l_y)\downarrow 2,\quad HH = (I \star h_x \star h_y)\downarrow 2,$$6 weather-guided gating (Liu et al., 3 Mar 2026).

The infrared-visible fusion paper “Exploring State Space Model in Wavelet Domain” introduces a “Wavelet-SSM module” and “cross-modal feature attention modulation,” but the provided paper text contains no technical content beyond metadata, so the exact terminology used in the paper, the architecture and hyperparameters of the WaveMamba Fusion Block, loss functions, complexity numbers, or comparison results cannot be confirmed from the extracted material (Zhang et al., 24 Mar 2025). This is important because it prevents the term WMFB from being retrospectively attributed to that paper on documentary grounds.

Taken together, these papers suggest a broader conceptual pattern: a “WaveMamba fusion block,” in an editorial rather than authorial sense, typically denotes a module that combines wavelet-domain decomposition with state-space modeling, uses task-specific rules to merge low- and high-frequency information, and often couples that fusion to an IDWT-based reconstruction path (Cai et al., 19 Dec 2025, Wang et al., 15 Aug 2025, Zhang et al., 5 Aug 2025, Ahmad et al., 2024, Liu et al., 3 Mar 2026). The exact implementation, however, varies sharply with task, dimensionality, and modality.

7. Conceptual significance and common misunderstandings

WMFB is best understood as a frequency-aware fusion operator rather than as a generic Mamba block. Its defining feature in the canonical WaveMamba detector is the explicit asymmetry between low- and high-frequency processing: low-frequency fusion is learned through Mamba-based selective scanning and gated interaction, whereas high-frequency fusion is delegated to an absolute-maximum selector (Zhu et al., 24 Jul 2025). This division encodes a domain assumption reported by the paper’s entropy analysis: infrared is more informative in low-frequency bands, while RGB is more informative in high-frequency bands (Zhu et al., 24 Jul 2025).

A second misunderstanding is that WMFB is synonymous with wavelet-domain fusion in infrared-visible image fusion. The available record does not support that equivalence. Some papers use wavelet-domain state-space modules for fusion, segmentation, enhancement, or restoration, but under names such as Wavelet-SSM, WHF, DAFF, HFEB, MDIF, or WSSB, and with different placement, loss design, and reconstruction strategy (Zhang et al., 24 Mar 2025, Cai et al., 19 Dec 2025, Wang et al., 15 Aug 2025, Zhang et al., 5 Aug 2025, Liu et al., 3 Mar 2026). WMFB is therefore both a specific module name and, only by plausible extension, a broader family resemblance.

A plausible implication is that the enduring contribution of WMFB lies less in a single fixed blueprint than in an architectural doctrine: decompose heterogeneous features into wavelet sub-bands, assign low-frequency semantics and high-frequency detail to different fusion mechanisms, and exploit state-space models where global context is most needed (Zhu et al., 24 Jul 2025, Cai et al., 19 Dec 2025, Liu et al., 3 Mar 2026). Under that reading, the original WaveMamba formulation remains the canonical reference, while later variants show how the same principle can be adapted to detection, segmentation, restoration, and hyperspectral classification.