ReDiffuse: Rotation Equivariant Diffusion Model for Multi-focus Image Fusion

Bo Li; Tingting Bao; Lingling Zhang; Weiping Fu; Yaxian Wang; Jun Liu

cs.CVMar 22, 2026

ReDiffuse: Rotation Equivariant Diffusion Model for Multi-focus Image Fusion

Authors: Bo Li, Tingting Bao, Lingling Zhang, Weiping Fu, Yaxian Wang, Jun Liu

Organizations: School of Computer Science and Technology, Xi’an Jiaotong University, Xi’an, Shaanxi, China · School of Information Engineering, Chang’an University, Xi’an, Shaanxi, China

Abstract

Diffusion models have achieved impressive performance on multi-focus image fusion (MFIF). However, a key challenge in applying diffusion models to the ill-posed MFIF problem is that defocus blur can make common symmetric geometric structures (e.g., textures and edges) appear warped and deformed, often leading to unexpected artifacts in the fused images. Therefore, embedding rotation equivariance into diffusion networks is essential, as it enables the fusion results to faithfully preserve the original orientation and structural consistency of geometric patterns underlying the input images. Motivated by this, we propose ReDiffuse, a rotation-equivariant diffusion model for MFIF. Specifically, we carefully construct the basic diffusion architectures to achieve end-to-end rotation equivariance. We also provide a rigorous theoretical analysis to evaluate its intrinsic equivariance error, demonstrating the validity of embedding equivariance structures. ReDiffuse is comprehensively evaluated against various MFIF methods across four datasets (Lytro, MFFW, MFI-WHU, and Road-MF). Results demonstrate that ReDiffuse achieves competitive performance, with improvements of 0.28-6.64% across six evaluation metrics. The code is available at https://github.com/MorvanLi/ReDiffuse.

Figures & tables

Figure 1. (a) is one defocused source image in MFIF containing typical geometric structures. (b) and (c) show fused results of the FusionDiff ( Li et al., 2024a ) and our proposed ReDiffuse , respectively. Even under defocus, rotation equivariance enables ReDiffuse to better preserve non-local directional similarity and local isotropic symmetry, providing clear visual evidence of its benefits for the MFIF task.

Figure 2. Fusion results generated by three diffusion fusion methods and our ReDiffuse . Our ReDiffuse achieves superior structural consistency and artifact suppression.

Figure 3. Four typical conventional regularization values for near-focus images from the Lytro dataset at different rotation angles. The circled regions show that many local structures at different orientations share similar geometric properties.

Figure 4. The MFIF framework based on the proposed ReDiffuse . Rot-E Conv denotes the B-Conv ( Xie et al., 2025 ) . Due to efficient parameter sharing, each Rot-E convolution uses only $\tfrac{1}{m}$ of the parameters of a regular convolution, where $m$ denotes the equivariant group and is set to 4. As a result, ReDiffuse is lightweight, reducing the number of parameters from 26.91M to 7.55M. Moreover, under the $m=4$ rotation group, ReDiffuse theoretically achieves zero equivariance error (see Corollary 3.5 ).

Figure 5. Illustration of the proposed theoretical results and their correspondence with the network modules.

Methods	Lytro Multi-focus Fusion Dataset						MFFW Multi-focus Fusion Dataset
Methods	Qabf $\uparrow$	QMI $\uparrow$	QG $\uparrow$	QP $\uparrow$	QE $\uparrow$	MS-SSIM $\uparrow$	Qabf $\uparrow$	QMI $\uparrow$	QG $\uparrow$	QP $\uparrow$	QE $\uparrow$	MS-SSIM $\uparrow$
MFF-GAN [InFus2021]	0.630	0.720	0.522	0.686	0.801	0.962	0.574	0.709	0.441	0.489	0.739	0.969
SwinFusion [JAS2022]	0.690	0.748	0.585	0.723	0.848	0.979	0.603	0.700	0.463	0.505	0.747	0.968
ZMFF [InfFus2023]	0.703	0.884	0.631	0.785	0.891	0.992	0.666	0.766	0.564	0.647	0.809	0.975
FusionDiff [ESWA2024]	0.711	0.903	0.640	0.795	0.856	0.991	0.659	0.821	0.574	0.654	0.750	0.980
CCF [NeurIPS2024]	0.491	0.799	0.411	0.587	0.476	0.949	0.450	0.761	0.375	0.432	0.391	0.944

Table 1. Main quantitative MFIF results. The best and second-best values are highlighted in red and blue , respectively.

Figure 6. Visual comparison for “MFFW-8” in MFFW dataset, with the demarcated areas zoomed in 5 times for easy observation.

Figure 7. Visualization of difference images. These images are obtained by subtracting the near-focus source image from the fused image. Less residual information indicates better fusion quality .

Figure 8. Illustration of rotation equivariant errors, defined as $\left|\Phi\!\left[\pi_{\tilde{R}}^{I}\right](I)-\pi_{\tilde{R}}^{F}\!\left[\Phi\right](I)\right|$ , where $\Phi$ denotes the original model (without Rot-E) or the proposed ReDiffuse model. Three images from the MFI-WHU dataset are evaluated under rotation angles of $90^{\circ}$ , $180^{\circ}$ , and $270^{\circ}$ . Please zoom for a better view.

Method	Mode	Qabf $\uparrow$	QMI $\uparrow$	QG $\uparrow$	QP $\uparrow$	QE $\uparrow$
FusionDiff	Original	0.66	0.82	0.57	0.65	0.75
FusionDiff	Proposed	0.69	0.84	0.60	0.69	0.78
VDMUFuse	Original	0.53	0.80	0.47	0.57	0.50
VDMUFuse	Proposed	0.57	0.82	0.53	0.61	0.54
Mask-DiFuser	Original	0.56	0.71	0.44	0.54	0.59
Mask-DiFuser	Proposed	0.59	0.73	0.48	0.56	0.61

Table 2. Generalization performance of the proposed rotation-equivariant method, evaluated on MFFW dataset.

	Configurations	Qabf $\uparrow$	QMI $\uparrow$	QG $\uparrow$	QP $\uparrow$	QE $\uparrow$
I	w/o rotation equivariance	0.67	0.81	0.59	0.67	0.76
II	G-Conv $\to$ B-Conv	0.68	0.81	0.60	0.69	0.81
III	F-Conv $\to$ B-Conv	0.70	0.84	0.61	0.71	0.80
	Ours	0.71	0.86	0.64	0.73	0.83

Table 3. Ablation study on the effectiveness and superiority of rotation equivariance, evaluated on the MFFW dataset.

Figure 9. Visualization of a failure case of ReDiffuse on the challenging MFFW dataset.

Explore similar work

Clarity Contrast and Similarity Selection for Multi-Focus Image Fusion
Aug 4, 2026Yicheng Zhang, Haoyou Deng, Zhiqiang Li +3Multi-Source Image Fusion
Key Laboratory of Image Processing and Intelligent Control, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, 430074, China
Degradation-Robust Fusion: An Efficient Degradation-Aware Diffusion Framework for Multimodal Image Fusion in Arbitrary Degradation Scenarios
Apr 10, 2026Yu Shi, Yu Liu, Zhong-Cheng Wu +3Multi-Source Image Fusion Diffusion Models
Department of Biomedical Engineering, Hefei University of Technology · Faculty of Information Engineering and Automation, Kunming University of Science and Technology · School of Information Science and Technology, University of Science and Technology of China
Inter-Image Pixel Shuffling for Multi-focus Image Fusion
Mar 7, 2026Huangxing Lin, Rongrong Ma, Cheng WangMulti-Source Image Fusion
College of Computer Science and Technology, Huaqiao University, Xiamen 361021, China

Aug 4, 2026J/K move · Enter open · S save

Clarity Contrast and Similarity Selection for Multi-Focus Image Fusion

Yicheng Zhang, Haoyou Deng, Zhiqiang Li, Wenti Yin+2

Key Laboratory of Image Processing and Intelligent Control, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, 430074, China

Multi-Source Image Fusion

Open paper AskPDF arXiv

Multi-focus image fusion (MFIF) aims to generate an all-in-focus image from multiple images of the same scene focused at different regions. Most existing deep learning-based methods lack explicit interaction between the source images, which limits their performance and interpretability. This paper presents a novel Clarity Contrast and Similarity Selection Network (CSNet), to bridge direct information exchange for MFIF. Specifically, by contrasting the clarity differences between source images within our proposed Clarity Contrast Attention Module (CCAM), we mutually enhance sharp features while suppressing blurry ones. This allows us to identify the exactly focused regions in each source and locate the focused-defocused boundaries. Moreover, the Defocus Spread Effect (DSE) degrades pixels in all source images around the boundaries. To further refine these ambiguous areas, we introduce a Similarity Selection Strategy, which reconstructs an initial clear image from source images and selects optimal pixels by comparing the similarity among them. Through this interactive approach, CSNet effectively preserves focused regions as well as recovering natural boundaries to fuse an all-in-focus output. Extensive experiments demonstrate that our method achieves state-of-the-art performance both quantitatively and qualitatively. Our code is available on Github: https://github.com/ZYC-HUST/CSNet.

Abstract

Explore similar work

Clarity Contrast and Similarity Selection for Multi-Focus Image Fusion

Degradation-Robust Fusion: An Efficient Degradation-Aware Diffusion Framework for Multimodal Image Fusion in Arbitrary Degradation Scenarios

Inter-Image Pixel Shuffling for Multi-focus Image Fusion