Wavelet-guided Cross-Domain Synthesis
- The paper introduces an algorithmic alignment module that recombines source targets with target backgrounds through wavelet filtering, yielding higher SNR and cleaner edges.
- Wavelet-guided synthesis leverages discrete wavelet transforms to separate low-frequency structure and high-frequency details, enabling explicit frequency control in cross-domain tasks.
- The approach circumvents expensive adversarial methods by operating directly in image space, supporting applications from infrared target detection to style augmentation.
Wavelet-guided Cross-domain Synthesis denotes, in the most explicit usage of the term, the Stage I alignment module of Ivan-ISTD, where source-domain small targets are recombined with target-domain backgrounds to generate training samples aligned with the target domain through wavelet filtering, structural matching, and seamless fusion (Li et al., 14 Oct 2025). More broadly, related work suggests a general synthesis pattern in which discrete wavelet transforms expose low-frequency structure and high-frequency detail as separate computational objects, enabling cross-domain generation, conditional diffusion, unpaired label-to-image translation, age progression, and style augmentation under domain shift (Friedrich et al., 2024, Yang et al., 17 Nov 2025, Eskandar et al., 2023, Li et al., 2018, Fu et al., 2022).
1. Problem formulation and conceptual scope
In Ivan-ISTD, Wavelet-guided Cross-domain Synthesis is situated entirely in the data space. Its goal is to take a source image crop containing a small target and an unlabeled target-domain image, then produce a new training image whose background comes from the target domain while the inserted target object comes from the source domain. This synthesis module is the first stage of a two-stage framework; the second stage performs Real-domain Noise Invariance Learning on the synthesized dataset, and no further adjustment to Stage I is required at inference time (Li et al., 14 Oct 2025).
This formulation addresses a specific failure mode of cross-domain infrared small target detection: the simultaneous presence of cross-domain shift and heteroscedastic noise perturbations. The Ivan-ISTD design treats background alignment as an overview problem rather than as adversarial feature alignment. The paper states that target-style backgrounds are transplanted into source training data “without requiring expensive adversarial style-transfer or feature-alignment at inference,” making the alignment mechanism explicitly constructive rather than implicit (Li et al., 14 Oct 2025).
A broader reading of the literature suggests that “cross-domain” in wavelet-guided synthesis need not refer only to source-versus-target datasets. In the surveyed works, the relevant domain gap may instead be defined by MR modality, semantic layout versus photograph, age condition, super-resolved prior versus diffused residual, or source-domain versus target-domain style statistics. This suggests that the unifying principle is not a single task definition, but the use of wavelet-domain factorization to control how information moves across heterogeneous representations.
2. Wavelet decomposition as the operative prior
The mathematical core of these methods is the discrete wavelet transform, which decomposes a signal into one low-frequency approximation and multiple high-frequency detail bands. In Ivan-ISTD, for a single-channel infrared image $X \in \mathbb{R}^{H \times W}$, one-level decomposition yields a low-pass component $L^{\mathrm{low}}$ and three detail components $H_k^{\mathrm{high}}$, followed by reconstruction with synthesis filters. The distinctive operation is edge-aware attenuation of the high-frequency bands using the local gradient magnitude of the low-frequency branch:
$\tilde H_k^{\mathrm{high}}(i,j) = H_k^{\mathrm{high}}(i,j)\times \frac{\left|\nabla L^{\mathrm{low}}(i,j)\right|} {\max_{x,y}\left|\nabla L^{\mathrm{low}}(x,y)\right|+\epsilon}, \quad k=1,2,3,$
with $\epsilon = 10^{-8}$. After filtering, the image is reconstructed as $F^{\mathrm{target}}=\mathrm{WaveletReconstruct}(L^{\mathrm{low}}, \{\tilde H_k^{\mathrm{high}}\})$, which the paper reports as having empirically higher SNR and cleaner edges (Li et al., 14 Oct 2025).
Other works use the same decomposition logic but change the dimensionality and downstream operator. In cWDM, a one-level fully three-dimensional DWT maps a volumetric MR scan into eight subbands, stacked into $x_0 \in \mathbb{R}^{8 \times \tfrac D2 \times \tfrac H2 \times \tfrac W2}$. The stated advantages are that the network processes tensors that are eight times smaller in memory and that the separated high-frequency subbands explicitly expose edges and fine detail, which diffusion models tend to reconstruct more faithfully than in pixel space (Friedrich et al., 2024).
In Wave-SAN, the 2D Haar DWT separates feature maps into $F_{\mathrm{low}}$ and $F_{\mathrm{high}}$, where low-frequency components are treated as carrying shape and style, and high-frequency components as carrying texture. The style augmentation module swaps only low-frequency statistics and preserves high-frequency bands unchanged before exact inverse reconstruction (Fu et al., 2022). In HDW-SR, Haar decomposition is used at every downsampling stage, with the low-frequency branch propagated and the high-frequency subbands cached as explicit guidance for diffusion-based restoration (Yang et al., 17 Nov 2025). Across these formulations, the common prior is that domain-relevant variation is more controllable when smooth structure and fine detail are disentangled.
3. The Ivan-ISTD alignment machine
The Ivan-ISTD synthesis module consists of three algorithmic submodules: Background Region Detection (BRD), Difficult-Target Selection (DIS), and Re-Generation via SSIM matching and Poisson fusion (Li et al., 14 Oct 2025).
BRD begins by decomposing a target-domain image, filtering the detail bands with the gradient-weighted rule above, and reconstructing $F^{\mathrm{target}}$. The reconstructed image is then partitioned into a uniform grid of non-overlapping subblocks $L^{\mathrm{low}}$0 of size $L^{\mathrm{low}}$1, with $L^{\mathrm{low}}$2. For $L^{\mathrm{low}}$3 images this yields $L^{\mathrm{low}}$4 and $L^{\mathrm{low}}$5. Each block is scored using edge density and Laplacian response:
$L^{\mathrm{low}}$6
$L^{\mathrm{low}}$7
followed by normalization to $L^{\mathrm{low}}$8 and averaging,
$L^{\mathrm{low}}$9
Blocks with $H_k^{\mathrm{high}}$0 are selected as likely background, with $H_k^{\mathrm{high}}$1. These are cropped to $H_k^{\mathrm{high}}$2 and bilinearly upsampled back to $H_k^{\mathrm{high}}$3 to obtain $H_k^{\mathrm{high}}$4 (Li et al., 14 Oct 2025).
DIS operates on the source domain during training. The current detector $H_k^{\mathrm{high}}$5 is run on a source crop $H_k^{\mathrm{high}}$6, and pixel accuracy and IoU are computed. If both fall below thresholds $H_k^{\mathrm{high}}$7 and $H_k^{\mathrm{high}}$8, the crop is added to the difficult-source pool $H_k^{\mathrm{high}}$9. The explicit criterion ensures that Stage I preferentially reuses hard examples rather than generic source targets (Li et al., 14 Oct 2025).
Re-Generation matches source targets to target backgrounds by structural similarity. For each source-target pair of patches, SSIM is computed; for each source patch, the method chooses among the top-$\tilde H_k^{\mathrm{high}}(i,j) = H_k^{\mathrm{high}}(i,j)\times \frac{\left|\nabla L^{\mathrm{low}}(i,j)\right|} {\max_{x,y}\left|\nabla L^{\mathrm{low}}(x,y)\right|+\epsilon}, \quad k=1,2,3,$0 candidates with $\tilde H_k^{\mathrm{high}}(i,j) = H_k^{\mathrm{high}}(i,j)\times \frac{\left|\nabla L^{\mathrm{low}}(i,j)\right|} {\max_{x,y}\left|\nabla L^{\mathrm{low}}(x,y)\right|+\epsilon}, \quad k=1,2,3,$1 whose SSIM is at least $\tilde H_k^{\mathrm{high}}(i,j) = H_k^{\mathrm{high}}(i,j)\times \frac{\left|\nabla L^{\mathrm{low}}(i,j)\right|} {\max_{x,y}\left|\nabla L^{\mathrm{low}}(x,y)\right|+\epsilon}, \quad k=1,2,3,$2, shuffles them, and uses the first unused candidate subject to $\tilde H_k^{\mathrm{high}}(i,j) = H_k^{\mathrm{high}}(i,j)\times \frac{\left|\nabla L^{\mathrm{low}}(i,j)\right|} {\max_{x,y}\left|\nabla L^{\mathrm{low}}(x,y)\right|+\epsilon}, \quad k=1,2,3,$3. Fusion is performed by solving the Poisson blending objective over the insertion region $\tilde H_k^{\mathrm{high}}(i,j) = H_k^{\mathrm{high}}(i,j)\times \frac{\left|\nabla L^{\mathrm{low}}(i,j)\right|} {\max_{x,y}\left|\nabla L^{\mathrm{low}}(x,y)\right|+\epsilon}, \quad k=1,2,3,$4, with solver tolerance approximately $\tilde H_k^{\mathrm{high}}(i,j) = H_k^{\mathrm{high}}(i,j)\times \frac{\left|\nabla L^{\mathrm{low}}(i,j)\right|} {\max_{x,y}\left|\nabla L^{\mathrm{low}}(x,y)\right|+\epsilon}, \quad k=1,2,3,$5 and 200 iterations. The resulting BRD $\tilde H_k^{\mathrm{high}}(i,j) = H_k^{\mathrm{high}}(i,j)\times \frac{\left|\nabla L^{\mathrm{low}}(i,j)\right|} {\max_{x,y}\left|\nabla L^{\mathrm{low}}(x,y)\right|+\epsilon}, \quad k=1,2,3,$6 DIS $\tilde H_k^{\mathrm{high}}(i,j) = H_k^{\mathrm{high}}(i,j)\times \frac{\left|\nabla L^{\mathrm{low}}(i,j)\right|} {\max_{x,y}\left|\nabla L^{\mathrm{low}}(x,y)\right|+\epsilon}, \quad k=1,2,3,$7 Re-Generation pipeline is described as the “wavelet-guided alignment machine,” whose purpose is to reduce the background-style gap between source and target domains by working directly in image space (Li et al., 14 Oct 2025).
4. Architectural variants across synthesis settings
| Work | Domain mapping | Wavelet-guided mechanism |
|---|---|---|
| Ivan-ISTD (Li et al., 14 Oct 2025) | source target crop + target background $\tilde H_k^{\mathrm{high}}(i,j) = H_k^{\mathrm{high}}(i,j)\times \frac{\left|\nabla L^{\mathrm{low}}(i,j)\right|} {\max_{x,y}\left|\nabla L^{\mathrm{low}}(x,y)\right|+\epsilon}, \quad k=1,2,3,$8 target-aligned training image | BRD, DIS, SSIM matching, Poisson fusion |
| cWDM (Friedrich et al., 2024) | three MR modalities $\tilde H_k^{\mathrm{high}}(i,j) = H_k^{\mathrm{high}}(i,j)\times \frac{\left|\nabla L^{\mathrm{low}}(i,j)\right|} {\max_{x,y}\left|\nabla L^{\mathrm{low}}(x,y)\right|+\epsilon}, \quad k=1,2,3,$9 missing modality | 3D DWT-conditioned DDPM on full volumes |
| HDW-SR (Yang et al., 17 Nov 2025) | pre-super-resolved prior + diffused residual $\epsilon = 10^{-8}$0 SR image | Haar-DWT down/up-sampling, sparse HF-guided cross-attention, DTB |
| USIS (Eskandar et al., 2023) | semantic layout $\epsilon = 10^{-8}$1 image without pairs | waveletSPADE generator, wavelet discriminator, segmentation cycle |
| WaveletGLCA-GAN (Li et al., 2018) | face + age label $\epsilon = 10^{-8}$2 aged face | predict LL/LH/HL/HH, fixed wavelet reconstruction |
| Wave-SAN (Fu et al., 2022) | episode style swap under domain shift | low-frequency swap, unchanged high-frequency, IDWT reconstruction |
In cWDM, the conditioning mechanism is deliberately simple. Each available MR modality is transformed by DWT, the resulting coefficients are stacked channel-wise, and those channels are concatenated to the noisy target-domain wavelets at every denoising step. No cross-attention or FiLM is used; the denoiser receives the concatenated tensor directly. The 3D U-Net has base channel width $\epsilon = 10^{-8}$3, four levels of down- and up-sampling, two residual 3D-convolution blocks per level, group normalization, SiLU activations, standard skip connections, and trilinear upsampling followed by $\epsilon = 10^{-8}$4 convolution in the decoder. By working in the wavelet domain, the paper reports approximately one-eighth of the memory footprint of a pixel-space alternative and nearly twice the speed when generating full $\epsilon = 10^{-8}$5 volumes end-to-end (Friedrich et al., 2024).
HDW-SR adopts a different synthesis decomposition. Diffusion is applied only to the residual map $\epsilon = 10^{-8}$6, not to the full image. Every encoder and diffusion-stage downsampling is replaced by Haar-DWT. Low-frequency channels from the noisy image form the query stream, while high-frequency channels from the pre-super-resolved image form keys and values in sparse cross-attention. A Dynamic Thresholding Block selects sparsity patterns by maximizing inter-class variance as in Otsu’s method. During reconstruction, inverse DWT is used level by level, and the paper states that the high-frequency wavelet guidance is never overwritten, but only used as guidance through cross-attention and DTB (Yang et al., 17 Nov 2025).
These variants show that wavelet-guided synthesis can be instantiated as direct image recomposition, conditional diffusion, or structured attention. The shared mechanism is explicit frequency-domain control over what is preserved, what is denoised, and what is newly synthesized.
5. Relation to unpaired translation, age progression, and style augmentation
Wavelet-guided synthesis also appears in unpaired label-to-image translation. In USIS, the core assumption is that a high-quality generated image should be segmented back to its semantic layout. The framework therefore combines a self-supervised segmentation loss with whole-image wavelet-based discrimination, and introduces a wavelet-domain generator called waveletSPADE. The generator synthesizes in wavelet space, uses waveletResBlocks with a waveletUpsample identity path, and applies pixelSPADE by temporarily reconstructing features to the spatial domain and mapping them back with DWT. The discriminator operates on $\epsilon = 10^{-8}$7 and $\epsilon = 10^{-8}$8, forcing explicit critique of LH, HL, and HH detail bands. On Cityscapes, ADE20K, and COCO-stuff, the reported USIS scores are $\epsilon = 10^{-8}$9, $F^{\mathrm{target}}=\mathrm{WaveletReconstruct}(L^{\mathrm{low}}, \{\tilde H_k^{\mathrm{high}}\})$0, and $F^{\mathrm{target}}=\mathrm{WaveletReconstruct}(L^{\mathrm{low}}, \{\tilde H_k^{\mathrm{high}}\})$1 in FID/mIoU respectively, compared with CycleGAN values $F^{\mathrm{target}}=\mathrm{WaveletReconstruct}(L^{\mathrm{low}}, \{\tilde H_k^{\mathrm{high}}\})$2, $F^{\mathrm{target}}=\mathrm{WaveletReconstruct}(L^{\mathrm{low}}, \{\tilde H_k^{\mathrm{high}}\})$3, and $F^{\mathrm{target}}=\mathrm{WaveletReconstruct}(L^{\mathrm{low}}, \{\tilde H_k^{\mathrm{high}}\})$4 (Eskandar et al., 2023).
In WaveletGLCA-GAN, wavelet guidance is integrated with explicit global-local decomposition for age synthesis. The generator comprises a Global-Specific Network, three Local-Specific Networks for forehead, eyes, and mouth, a feature-fusion network, a Wavelet Prediction Network with four parallel subband branches, and a fixed wavelet reconstruction layer implementing inverse DWT filters. The network predicts coefficient maps rather than directly generating the output image. The optimization includes adversarial, identity preserving, age preserving, pixel-wise, and total variation losses. Reported ablations show that WaveletGLCA-GAN improves age-estimation fidelity relative to image-domain GLCA-GAN and a no-local Wavelet-GAN baseline; on CACD2000, mean predicted ages for the $F^{\mathrm{target}}=\mathrm{WaveletReconstruct}(L^{\mathrm{low}}, \{\tilde H_k^{\mathrm{high}}\})$5, $F^{\mathrm{target}}=\mathrm{WaveletReconstruct}(L^{\mathrm{low}}, \{\tilde H_k^{\mathrm{high}}\})$6, and $F^{\mathrm{target}}=\mathrm{WaveletReconstruct}(L^{\mathrm{low}}, \{\tilde H_k^{\mathrm{high}}\})$7 groups are $F^{\mathrm{target}}=\mathrm{WaveletReconstruct}(L^{\mathrm{low}}, \{\tilde H_k^{\mathrm{high}}\})$8, $F^{\mathrm{target}}=\mathrm{WaveletReconstruct}(L^{\mathrm{low}}, \{\tilde H_k^{\mathrm{high}}\})$9, and $x_0 \in \mathbb{R}^{8 \times \tfrac D2 \times \tfrac H2 \times \tfrac W2}$0, and face verification on CACD2000 reaches $x_0 \in \mathbb{R}^{8 \times \tfrac D2 \times \tfrac H2 \times \tfrac W2}$1, $x_0 \in \mathbb{R}^{8 \times \tfrac D2 \times \tfrac H2 \times \tfrac W2}$2, and $x_0 \in \mathbb{R}^{8 \times \tfrac D2 \times \tfrac H2 \times \tfrac W2}$3 TAR@FAR$x_0 \in \mathbb{R}^{8 \times \tfrac D2 \times \tfrac H2 \times \tfrac W2}$4 for Test$x_0 \in \mathbb{R}^{8 \times \tfrac D2 \times \tfrac H2 \times \tfrac W2}$5AG1, Test$x_0 \in \mathbb{R}^{8 \times \tfrac D2 \times \tfrac H2 \times \tfrac W2}$6AG2, and Test$x_0 \in \mathbb{R}^{8 \times \tfrac D2 \times \tfrac H2 \times \tfrac W2}$7AG3 (Li et al., 2018).
Wave-SAN moves the same principle into episodic feature space for cross-domain few-shot learning. StyleAug performs an AdaIN-style swap only on low-frequency subbands, reconstructing augmented features with unchanged high-frequency components. A symmetric KL self-supervised loss enforces consistency between predictions on original and style-augmented episodes. DWT/IDWT and StyleAug are inserted after ResNet-10 blocks 1, 2, and 3, but are bypassed entirely at test time. On the FWT benchmark in 5-way 1-shot evaluation averaged over 1,000 episodes, the reported GNN baseline improves from $x_0 \in \mathbb{R}^{8 \times \tfrac D2 \times \tfrac H2 \times \tfrac W2}$8 to $x_0 \in \mathbb{R}^{8 \times \tfrac D2 \times \tfrac H2 \times \tfrac W2}$9 on CUB, $F_{\mathrm{low}}$0 to $F_{\mathrm{low}}$1 on Cars, $F_{\mathrm{low}}$2 to $F_{\mathrm{low}}$3 on Places, and $F_{\mathrm{low}}$4 to $F_{\mathrm{low}}$5 on Plantae (Fu et al., 2022).
Taken together, these works suggest that wavelet guidance is not tied to a single training regime. It can supervise unpaired generation through a discriminator, mediate conditional generation through coefficient prediction, or span style distributions through parameter-free feature recomposition.
6. Empirical behavior, scope, and common misconceptions
A recurring empirical pattern is that wavelet guidance is used to avoid coarse approximations that suppress fine structure. In cWDM, four separate models are trained for missing-modality synthesis on the BraTS 2024 challenge data, which contains 1,251 train and 219 validation subjects with co-registered T1, T1ce, T2, and FLAIR volumes at $F_{\mathrm{low}}$6. Intensities are clipped to the $F_{\mathrm{low}}$7–$F_{\mathrm{low}}$8 percentiles and normalized to $F_{\mathrm{low}}$9. Training uses 1.2 M steps, Adam with learning rate $F_{\mathrm{high}}$0, batch size 1, and a diffusion schedule of $F_{\mathrm{high}}$1 with linear $F_{\mathrm{high}}$2 from $F_{\mathrm{high}}$3 to $F_{\mathrm{high}}$4. Reported validation performance includes $F_{\mathrm{high}}$5, $F_{\mathrm{high}}$6 dB, and $F_{\mathrm{high}}$7 when generating T1 from $F_{\mathrm{high}}$8; $F_{\mathrm{high}}$9, $F^{\mathrm{target}}$0, and $F^{\mathrm{target}}$1 for FLAIR synthesis; and overall random-drop performance of $F^{\mathrm{target}}$2, $F^{\mathrm{target}}$3, $F^{\mathrm{target}}$4 (Friedrich et al., 2024).
Several misconceptions are directly contradicted by the literature. First, wavelet-guided synthesis is not synonymous with learned style transfer. Ivan-ISTD Stage I is purely algorithmic and has no trainable weights; its role is to recompose data before the detector learns anything from the aligned dataset (Li et al., 14 Oct 2025). Second, wavelet guidance is not restricted to 2D image translation. cWDM uses a fully three-dimensional wavelet representation and performs diffusion on full volumetric scans rather than on slices or patches (Friedrich et al., 2024). Third, wavelet guidance does not require high-frequency channels to be directly rewritten by the generator. In HDW-SR, the high-frequency bands of the guidance image are used through sparse cross-attention plus DTB and are “never overwritten,” while diffusion acts on low-frequency structure and residual restoration (Yang et al., 17 Nov 2025).
The practical scope of these methods is broader than the individual benchmark settings. cWDM explicitly states that the same framework can be retrained for CT $F^{\mathrm{target}}$5 MR, MR $F^{\mathrm{target}}$6 PET, and mask-conditioned anatomically guided image generation, and that any anatomically aligned paired volumes can be handled by the same DWT–diffusion–IDWT pipeline (Friedrich et al., 2024). HDW-SR states that its modular structure “lends itself to straightforward re-use in other diffusion-based restoration tasks” (Yang et al., 17 Nov 2025). These claims suggest that wavelet-guided cross-domain synthesis is best understood as a reusable design principle: separate frequency bands, allocate cross-domain guidance to the appropriate bands, and reconstruct with an exactly invertible operator whenever preservation of detail is the central constraint.