UNet-AF: Alias-free UNet architectures

Jérémy Scanvic; Quentin Barthélemy; Julián Tachella

cs.CVMar 11, 2026

UNet-AF: Alias-free UNet architectures

Authors: Jérémy Scanvic, Quentin Barthélemy, Julián Tachella

Organizations: ENSL, CNRS, Laboratoire de Physique, F-69364 Lyon, France. · Hirsch Secure, Lyon, France.

Abstract

The simplicity and effectiveness of UNet architectures make them ubiquitous in image restoration, segmentation, and diffusion models. They are often assumed to be equivariant to translations, yet they traditionally consist of layers that are known to be prone to aliasing, which hinders their equivariance in practice. To overcome this limitation, we show how to build sub-pixel translation-equivariant UNet architectures by appropriately choosing their main components (convolution, pooling, downsampling, activation, and normalization layers) to be alias-free. We evaluate the proposed equivariant architectures against non-equivariant baselines on image restoration tasks and observe competitive performance with a significant increase in measured equivariance. Through extensive ablation studies, we also demonstrate the importance of every architectural choice to achieve high equivariance. Our implementation is available at https://github.com/jscanvic/UNet-AF

Figures & tables

Fig. 1: Equivariance and robustness to circular translations. (Top) Performance on an horizontally translated image, with 0.01 px steps, (bottom) adversarial performance on the worst translation up to a maximum displacement, with 0.25 px steps. UNet-AF is stable and robust, unlike the baselines.

Fig. 2: Overview of the architecture. UNet-AF is obtained by changing the architectures of [ 1 ] , and [ 2 ] . Anti-aliasing filters are added in upsampling layers, max-pooling layers are replaced with blur pooling layers, ReLU activations are replaced with filtered GELU layers and batch normalization is replaced by alias-free layer normalization. The padding mode and residual connection are specified on a per-task basis.

Method	Difference	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	EQUIV $\uparrow$	STAB $\uparrow$	FPS $\uparrow$
UNet-Ronneberger		31.28 ± 2.13	0.9316 ± 0.0271	0.0377 ± 0.0235	37.60 ± 2.62	9.10	2494
UNet-Jin	+ Res conn.	31.64 ± 2.51	0.9280 ± 0.0345	0.0402 ± 0.0254	38.20 ± 2.22	12.86	2114
	+ LayerNorm-AF	32.16 ± 2.37	0.9383 ± 0.0211	0.0355 ± 0.0210	38.41 ± 1.99	18.40	1230
	+ Circular pad.	32.46 ± 2.45	0.9417 ± 0.0209	0.0323 ± 0.0199	43.16 ± 2.38	20.63	982
	+ Filtered GELU	32.22 ± 2.39	0.9383 ± 0.0201	0.0331 ± 0.0193	43.55 ± 2.77	20.48	302
	+ Filtered Upconv.	32.24 ± 2.35	0.9382 ± 0.0201	0.0333 ± 0.0197	46.33 ± 2.59	20.34	293

TABLE I: Performance on circular deblurring. The changes in Section II lead to a steady increase in equivariance (EQUIV), stabilize the training dynamics and improve performance (PSNR, SSIM, LPIPS), at the expense of a higher computational cost (FPS). Best metric in bold , second best underlined . Values: avg ± st.d.

Fig. 3: Evaluation PSNR throughout training for circular deblurring. Our proposed model has significantly smoother training dynamics than the non-equivariant baselines.

Replacement	Variant	PSNR $\uparrow$	EQUIV $\uparrow$
Res. connection	Present	32.22 ± 2.32	81.63 ± 0.26
	Absent	31.96 ± 2.33	77.77 ± 2.32
Normalization	LayerNorm-AF	32.22 ± 2.32	81.63 ± 0.26
	BatchNorm	32.21 ± 2.31	80.62 ± 8.14
	InstanceNorm	31.50 ± 2.19	78.40 ± 0.42
	LayerNorm	32.31 ± 2.36	54.19 ± 3.03

TABLE II: Ablation study. Each design choice made in Section II is explored in terms of performance and equivariance. Best metric in bold , second best underlined . Values: avg ± st.d.

Fig. 4: Reconstructions of a circularly blurred image. The PSNR is displayed in the bottom-left corner of each image.

Method	Task	PSNR $\uparrow$	EQUIV $\uparrow$
UNet-Ronneberger	Valid deblurring	31.67 ± 2.33	39.96 ± 2.71
UNet-Jin		32.25 ± 2.41	42.82 ± 2.96
UNet-AF		32.30 ± 2.42	46.57 ± 2.58
Blurred images		26.85 ± 2.55	N/A
UNet-Ronneberger	Gaussian denoising	29.70 ± 1.56	39.80 ± 1.31
UNet-Jin		29.58 ± 1.61	40.64 ± 1.78

TABLE III: Performance on additional tasks. We evaluate UNet-AF on valid deblurring, and Gaussian denoising. Best metric in bold , second best underlined . Values: avg ± st.d.

Filters	Padding	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	EQUIV $\uparrow$	STAB $\uparrow$	FPS $\uparrow$
Kaiser	Zeros	32.41 ± 2.42	0.9404 ± 0.0209	0.0363 ± 0.0210	43.12 ± 1.89	22.30	556
Kaiser	Circular	26.79 ± 2.24	0.8315 ± 0.0367	0.2247 ± 0.0806	89.35 ± 2.35	13.45	571
Tuned	Zeros	32.45 ± 2.44	0.9413 ± 0.0211	0.0362 ± 0.0210	43.94 ± 2.06	23.29	534
Tuned	Circular	32.39 ± 2.40	0.9408 ± 0.0207	0.0307 ± 0.0181	49.36 ± 2.84	23.29	552

TABLE IV: Ablation on UNet-AF Fast. Best metric in bold , second best underlined . Values: avg ± st.d.

Explore similar work

Fast Equivariant Imaging: Accelerating Unsupervised Learning and Model Adaptation via Inexact Splitting
Jul 9, 2025Guixian Xu, Jinglai Li, Junqi TangImage Reconstruction Computational Imaging
School of Mathematics, University of Birmingham, Birmingham B15 2TT, United Kingdom
Aligning Network Equivariance with Data Symmetry: A Theoretical Framework and Adaptive Approach for Image Restoration
May 13, 2026Feiyu Tan, Qi Xie, Zongben Xu +1Equivariant Neural Networks Image Restoration
School of Mathematics and Statistics, Xi’an Jiaotong University, Xi’an, Shaanxi, China
Normalization Equivariance for Arbitrary Backbones, with Application to Image Denoising
May 5, 2026Youssef Saied, François FleuretImage Corruption Robustness Neural Network Robustness
Department of Computer Science, University of Geneva, Switzerland · Meta FAIR.

Jul 9, 2025J/K move · Enter open · S save

Fast Equivariant Imaging: Accelerating Unsupervised Learning and Model Adaptation via Inexact Splitting

Guixian Xu, Jinglai Li, Junqi Tang

School of Mathematics, University of Birmingham, Birmingham B15 2TT, United Kingdom

Image Reconstruction Computational Imaging

Open paper AskPDF arXiv

In this work, we propose Fast Equivariant Imaging (FEI), a novel unsupervised learning framework to rapidly and efficiently train deep imaging networks without ground-truth data. FEI reformulates the EI objective through an inexact variable-splitting scheme, decoupling network training from an auxiliary restoration step implemented with a plug-and-play denoiser, this novel unsupervised scheme shows superior efficiency and performance compared to the standard Equivariant Imaging paradigm. In particular, our FEI schemes achieve an order-of-magnitude (10x) acceleration over standard EI on training U-Net for X-ray CT reconstruction and image inpainting, with improved generalization performance. Beyond offline training, the proposed scheme also enables efficient test-time adaptation of a pretrained model to individual samples, to secure further performance improvements. Extensive experiments show that the proposed approach provides a noticeable efficiency and performance gain over existing unsupervised methods and model adaptation techniques.

Abstract

Explore similar work

Fast Equivariant Imaging: Accelerating Unsupervised Learning and Model Adaptation via Inexact Splitting

Aligning Network Equivariance with Data Symmetry: A Theoretical Framework and Adaptive Approach for Image Restoration

Normalization Equivariance for Arbitrary Backbones, with Application to Image Denoising