WaFusion: Wavelet-Enhanced Face Morphing

• WaFusion is a face morph generation framework that uses Haar wavelet decomposition to isolate the low-frequency (LL) sub-band for diffusion-based morph synthesis.
• It preserves high-frequency details by analytically averaging corresponding sub-bands, reducing artifacts and computational cost while maintaining texture consistency.
• Evaluated on multiple biometric datasets, WaFusion demonstrates robust performance with strong metrics such as EER and AUC, making it valuable for attack modeling and defense evaluation.

Searching arXiv for the WaFusion paper and closely related wavelet-based fusion/morphing work. I’m checking arXiv records relevant to WaFusion and adjacent wavelet-based fusion methods. WaFusion is a face morph generation framework that combines wavelet decomposition with diffusion-based generation to synthesize high-quality morphed face images for biometric security research. It is presented as a response to face morphing attacks, in which two identities are blended into a single facial image that can resemble both people closely enough to pass identity verification; the stated application domain includes passport issuance and automated border control. The framework processes only the low-low wavelet sub-band with a diffusion autoencoder while preserving high-frequency sub-bands by averaging them from the two source images, and then reconstructs the final morph by inverse wavelet transform. In the reported experiments on FERET, FRGC, FRLL, and WVU Twin, it is described as producing high-resolution morphs with fewer artifacts and strong performance under biometric evaluation metrics including APCER, BPCER, EER, and AUC (Hosseini et al., 15 Jul 2025).

1. Problem setting and security context

WaFusion is situated in the literature on biometric face morphing attacks. In this setting, two bona fide face images are blended into a single image that can match both contributors closely enough to undermine identity verification systems. The stated motivation is not limited to image synthesis quality: stronger morph generation is also used to stress-test verification and morph-detection systems, since realistic attacks reveal weaknesses in biometric pipelines (Hosseini et al., 15 Jul 2025).

The framework is therefore best understood as a morph-generation method for attack modeling and defense evaluation, rather than as a generic image-fusion architecture. A common source of confusion is the term “fusion” itself. In WaFusion, “fusion” refers to combining facial identity content from two source images in a way that yields a plausible morphed face; it does not denote the broader multimodal fusion problem addressed by infrared-visible, RGB-NIR, medical, radar-camera, or RGB-IR detection systems. This distinction is important because the biometric objective is not merely visual pleasantness, but the generation of attack-capable samples that remain realistic and hard to detect.

The problem formulation in WaFusion assumes two aligned source images, resized to $512 \times 512$, and then decomposed into wavelet sub-bands. The operational claim is that global facial structure is concentrated in the low-frequency approximation, whereas edges and textures are concentrated in the higher-frequency sub-bands. This motivates selective generative processing: semantic morphing is concentrated in the low-frequency pathway, while detail retention is handled analytically rather than by full-image synthesis.

2. Core representation: wavelet decomposition and selective generation

The defining design decision in WaFusion is to apply a single-level Haar discrete wavelet transform to each aligned input face. The decomposition produces four sub-bands: $LL$, $LH$, $HL$, and $HH$. The paper characterizes these respectively as low-frequency approximation, horizontal-edge detail, vertical-edge detail, and diagonal/high-frequency texture detail (Hosseini et al., 15 Jul 2025).

The Haar decomposition is written as

$$X_{pq} = F_p^\top \cdot X \cdot F_q,\quad p,q \in \{L,H\},$$

where $X \in \mathbb{R}^{H \times W}$, yielding four sub-bands of size $\frac{H}{2} \times \frac{W}{2}$. Within this representation, the $LL$ band is treated as the carrier of coarse semantic facial structure, while $LH$, $LL$0, and $LL$1 are treated as carriers of sharp detail and local texture.

Only the $LL$2 sub-band is passed to the morph-generation block. The three high-frequency sub-bands are not synthesized by the diffusion model; instead, corresponding bands from the two source images are averaged. The reconstructed morph is obtained through inverse wavelet transform. The stated rationale is twofold. First, processing only the $LL$3 band reduces computation because the diffusion model is not applied to all wavelet components. Second, preserving the high-frequency bands analytically helps retain edge and texture consistency while reducing visible artifacts. The ablation study explicitly supports this design choice: on the Twin dataset, “WaFusion (All sub-bands)” yields LPIPS $LL$4, SSIM $LL$5, whereas “WaFusion (LL sub-band)” yields LPIPS $LL$6, SSIM $LL$7; the paper also states that processing all four bands would cost roughly $LL$8 more computation (Hosseini et al., 15 Jul 2025).

This selective wavelet-domain design places WaFusion in a specific technical niche. It is neither a purely landmark-based morphing system nor a pure full-image diffusion pipeline. Instead, it uses wavelets to decouple global structural morph synthesis from high-frequency detail preservation.

3. Diffusion autoencoder, interpolation, and reconstruction pipeline

The low-frequency morph-generation block is implemented with Diffusion Autoencoders and sampled using DDIM. The forward diffusion process is described by

$LL$9

and equivalently

$LH$0

with

$LH$1

The reverse process is written as

$LH$2

with the denoising function $LH$3 implemented using a U-Net, and the implementation uses 100 DDIM steps (Hosseini et al., 15 Jul 2025).

The diffusion autoencoder itself is described as a dual-encoder design. One encoder is semantic and captures structural or identity-related content; the other is stochastic and captures details such as textures, hair direction, and clothing. The two latent streams are interpolated differently. Image-space interpolation is given by

$LH$4

while latent interpolation is given by

$LH$5

Semantic features are linearly interpolated, while stochastic features use spherical linear interpolation. The stated intent is to preserve facial landmarks while producing a smooth blend of appearance details (Hosseini et al., 15 Jul 2025).

From a systems perspective, the full WaFusion pipeline is therefore a hybrid of three operations: alignment and resizing to $LH$6; Haar-wavelet decomposition with $LH$7-only generative processing; and inverse wavelet reconstruction using the generated $LH$8 band together with averaged $LH$9, $HL$0, and $HL$1 bands. This suggests a targeted allocation of generative capacity: model the component that carries most coarse identity structure, and preserve the remainder through deterministic signal-domain combination.

4. Datasets, evaluation protocol, and reported empirical results

WaFusion is evaluated on four datasets: FERET, FRGC, FRLL, and WVU Twin. The paper states that the verifier used for evaluation is FaceNet with an Inception backbone pretrained on VGGFace2, and that biometric metrics are evaluated following the ISO/IEC 30107-3 framework. The reported metrics are APCER, BPCER, EER, and AUC (Hosseini et al., 15 Jul 2025).

Dataset	Characteristics	Generated morphs
WVU Twin	2,268 unique identities; frontal, neutral, plain backgrounds; resolutions $HL$2 to $HL$3	2,971
FRLL	102 identities; passport-style frontal images; $HL$4	1,122
FRGC	subset of 765 identities; $HL$5	964
FERET	1,199 identities; $HL$6	529

The biometric findings are reported as follows. On FRGC, WaFusion obtains AUC $HL$7, EER $HL$8, and APCER@BPCER $HL$9 $HH$0. On FERET, it obtains AUC $HH$1, EER $HH$2, and APCER@BPCER $HH$3 $HH$4. On FRLL, StyleGAN is stated to be slightly stronger on some metrics numerically, but WaFusion is still reported at AUC $HH$5 and EER $HH$6. On WVU Twin, WaFusion reaches AUC $HH$7, EER $HH$8, and APCER@BPCER $HH$9 $$X_{pq} = F_p^\top \cdot X \cdot F_q,\quad p,q \in \{L,H\},$$0 (Hosseini et al., 15 Jul 2025).

Qualitatively, the paper contrasts WaFusion with OpenCV morphing, FaceMorpher, StyleGAN, a Diffusion Autoencoders baseline, and an LL-only baseline. OpenCV and FaceMorpher are reported to produce visible artifacts, especially around contours and hairlines. StyleGAN is described as visually pleasing in some cases but sometimes introducing GAN-specific inconsistencies. The pure diffusion baseline is described as strong in facial structure but limited to $$X_{pq} = F_p^\top \cdot X \cdot F_q,\quad p,q \in \{L,H\},$$1, whereas WaFusion produces $$X_{pq} = F_p^\top \cdot X \cdot F_q,\quad p,q \in \{L,H\},$$2 morphs. The claimed overall pattern is sharper and more realistic morphs with fewer artifacts.

These results should be read in the specific biometric sense used by the paper. The framework is evaluated not only by visual quality but also by how difficult the generated morphs are for the FaceNet-based verifier and associated presentation-attack evaluation protocol.

5. Position within wavelet-based fusion and generation research

WaFusion belongs to a broader research direction that combines learned models with wavelet-domain structure/detail separation, but its application and mechanism are more specific than those of general image-fusion systems. Earlier work such as “WaveFuse: A Unified Deep Framework for Image Fusion with Discrete Wavelet Transform” combines a deep encoder-decoder with a DWT-based fusion module for multi-exposure fusion, multi-modal medical fusion, multi-focus fusion, and infrared-visible fusion; its key claim is that learned feature maps should not be fused only by averaging or concatenation, but should first be decomposed into low- and high-frequency components and then fused with adaptive rules such as regional energy and variance-based selection (Liu et al., 2020).

More recent multimodal frameworks extend the same frequency-aware logic in different directions. “WIFE-Fusion: Wavelet-aware Intra-inter Frequency Enhancement for Multi-model Image Fusion” introduces Intra-Frequency Self-Attention and Inter-Frequency Interaction to model same-band cross-modal complementarity and cross-frequency residual transfer in tasks including IVIF, RGBNIR, and medical image fusion (Zhang et al., 4 Jun 2025). “WaveMamba: Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection” uses DWT/IDWT to separate low-frequency structure from high-frequency detail, applying Mamba-based low-frequency fusion and absolute-maximum high-frequency fusion for RGB-IR object detection (Zhu et al., 24 Jul 2025). A related IVIF paper, “Exploring State Space Model in Wavelet Domain: An Infrared and Visible Image Fusion Network via Wavelet Transform and State Space Model,” proposes Wavelet-Mamba, which integrates wavelet transform with an SSM through a Wavelet-SSM module and cross-modal feature attention modulation (Zhang et al., 24 Mar 2025).

Within that landscape, WaFusion differs in two decisive respects. First, its target is face morph generation rather than generic multimodal fusion or downstream detection. Second, its wavelet decomposition is used to offload only the $$X_{pq} = F_p^\top \cdot X \cdot F_q,\quad p,q \in \{L,H\},$$3 band to a diffusion autoencoder while reconstructing high-frequency content through band averaging. This suggests a narrower but technically well-defined objective: concentrate generative modeling where coarse facial identity structure is dominant, and preserve the remainder through inverse wavelet synthesis.

6. Interpretation, limitations, and nomenclature

The main interpretation advanced by the paper is that wavelets and diffusion are complementary. Wavelets isolate the structural part of the image and reduce spatial size; diffusion provides realistic generative modeling; and processing only the $$X_{pq} = F_p^\top \cdot X \cdot F_q,\quad p,q \in \{L,H\},$$4 band is intended to improve efficiency while preserving fine detail through the untouched high-frequency bands. A plausible implication is that WaFusion treats wavelet decomposition not merely as preprocessing, but as an architectural prior about where identity-relevant morph content resides.

The paper’s explicitly stated future directions are to move beyond single-level Haar decomposition, explore other wavelet bases such as Daubechies, selectively process some high-frequency sub-bands at deeper levels, and extend the framework to real-time video-based biometric authentication (Hosseini et al., 15 Jul 2025). These proposals indicate that the current design is intentionally simple in its wavelet choice and depth.

Nomenclature around “WaFusion” is not fully stable across the surrounding literature. In the biometric context, WaFusion denotes the wavelet-enhanced diffusion framework for face morph generation (Hosseini et al., 15 Jul 2025). In a different domain, “Wave-Supported Hybrid Fast-Thermal p-$$X_{pq} = F_p^\top \cdot X \cdot F_q,\quad p,q \in \{L,H\},$$5B Fusion” is described as directly related to “WaFusion” if that term is intended as shorthand for wave-assisted fusion, where waves channel $$X_{pq} = F_p^\top \cdot X \cdot F_q,\quad p,q \in \{L,H\},$$6-particle energy to fast protons in a hybrid fast-thermal plasma (Kolmes et al., 2022). Separately, a 2026 radar-camera maritime detection framework is named PhysFusion, with a note that the query term “WaFusion” appears to refer to that system, although the paper itself consistently uses “PhysFusion” (Wan et al., 2 Mar 2026). For technical precision, WaFusion in current arXiv usage most directly identifies the face morph generation framework, while the other usages are domain-specific ambiguities rather than the same method.

In summary, WaFusion is a wavelet-enhanced diffusion framework for face morph generation that decomposes aligned faces into Haar sub-bands, synthesizes only the low-frequency structural component with a diffusion autoencoder and DDIM, preserves high-frequency detail by band averaging, and reconstructs $$X_{pq} = F_p^\top \cdot X \cdot F_q,\quad p,q \in \{L,H\},$$7 morphs through inverse wavelet transform. Its significance lies in the combination of biometric attack modeling, wavelet-domain structure/detail separation, and selective generative processing, which together define a specific methodological variant within the broader family of wavelet-aware fusion and generation systems.