Towards Universal Computational Aberration Correction in Photographic Cameras: A Comprehensive Benchmark Analysis

Xiaolong Qian; Qi Jiang; Yao Gao; Lei Sun; Zhonghua Yi; Kailun Yang; Luc Van Gool; Kaiwei Wang

cs.CVMar 12, 2026

Towards Universal Computational Aberration Correction in Photographic Cameras: A Comprehensive Benchmark Analysis

Authors: Xiaolong Qian, Qi Jiang, Yao Gao, Lei Sun, Zhonghua Yi, Kailun Yang, Luc Van Gool, Kaiwei Wang

Organizations: Zhejiang University · INSAIT, Sofia University “St. Kliment Ohridski” · Hunan University

Abstract

Prevalent Computational Aberration Correction (CAC) methods are typically tailored to specific optical systems, leading to poor generalization and labor-intensive re-training for new lenses. Developing CAC paradigms capable of generalizing across diverse photographic lenses offers a promising solution to these challenges. However, efforts to achieve such cross-lens universality within consumer photography are still in their early stages due to the lack of a comprehensive benchmark that encompasses a sufficiently wide range of optical aberrations. Furthermore, it remains unclear which specific factors influence existing CAC methods and how these factors affect their performance. In this paper, we present comprehensive experiments and evaluations involving 24 image restoration and CAC algorithms, utilizing our newly proposed UniCAC, a large-scale benchmark for photographic cameras constructed via automatic optical design. The Optical Degradation Evaluator (ODE) is introduced as a novel framework to objectively assess the difficulty of CAC tasks, offering credible quantification of optical aberrations and enabling reliable evaluation. Drawing on our comparative analysis, we identify three key factors -- prior utilization, network architecture, and training strategy -- that most significantly influence CAC performance, and further investigate their respective effects. We believe that our benchmark, dataset, and observations contribute foundational insights to related areas and lay the groundwork for future investigations. Benchmarks, codes, and Zemax files will be available at https://github.com/XiaolongQian/UniCAC.

Figures & tables

Figure 1: Comparison of the linear relationship between RMS radius and ODE with the CAC results. (a) We select $12$ lenses and use the same model [ 2 ] for lens-specific training. The red line shows the least squares linear fit, with $R^{2}$ indicating correlation strength (closer to $1$ means stronger correlation). (b) Visual results: Lens1, with a larger RMS radius, retains more details, while Lens2, with a smaller RMS radius, loses fine structures.

Figure 2: Optical Degradation Evaluator (ODE) pipeline. It assesses lens optical degradation severity via standard checkerboard imaging. The spatial branch calculates OIQ for various FoVs, the channel branch for different channels.

Figure 3: Benchmark overview. The sampled lenses are divided into $5$ levels based on ODE, with $3$ lenses from each level showcased to facilitate presentation. Each lens is illustrated with the ray tracing diagram, PSFs for $3$ FoVs, and corresponding aberration image patches. The distribution of quantification dimensions for each lens is also presented in a bar chart.

Figure 4: (a) The ray tracing diagram, spot diagrams and RMS spot sizes produced by the simulation framework closely resemble those obtained from Zemax . The tested system is a single-lens optical system. (b) Comparison between simulation and real-shot aberration images.

Figure 5: (a) Relevance between ODE components and CAC performance. (b) Distribution of different quantification dimensions in our sampled lens library.

Figure 6: (a) Weighted average value of PSNR, SSIM, and OIQE. (b) Weighted average value of Image Fidelity, Optical Quality, and Perceptual Quality. (c) Effect of varying coefficient ratios of $OIQ$ and $U_{s}$ on the linearity of ODE and O. P. (d) Effect of varying coefficient ratios of $U_{c}$ on the linearity of ODE and O. P.

Model		Blind?	Model Type	Learning Paradigm
CAC	WienerDeconv [MITPress49] [ 49 ]	✗	Optimization	-
	FSANet [OE22] [ 27 ]	✗	Optimization&CNN	Regression
	PART [TIP24] [ 19 ]	✗	Transformer	Regression
	Peng et al . [TOG19] [ 34 ]	✓	CNN	GAN
	DFUnet [TOG21] [ 5 ]	✓	CNN	Regression
	FOV-KPN [ICCV21] [ 4 ]	✓	CNN	Regression

Table 1: Overview of the

24

evaluated methods.

Method	Params (M)	FLOPs (G)	Time (ms)	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	OIQE $\uparrow$	ClipIQA $\uparrow$	O. P. $\uparrow$
Wiener [MITPress49] [ 49 ]			0.21	19.62	0.699	0.448	122.8	0.475	0.242	0.972 (24)
FSANet [OE22] [ 27 ]	12.86	54.14	7.57	21.60	0.751	0.268	88.19	0.298	0.344	1.191 (20)
PART [TIP24] [ 19 ]	19.35	51.86	79.09	28.10	0.866	0.228	43.66	0.608	0.389	1.494 (7)
Peng et al . [TOG19] [ 34 ]	44.59	21.68	1.01	21.81	0.744	0.318	104.5	0.279	0.267	1.108 (21)
DFUnet [TOG21] [ 5 ]	4.23	18.90	5.99	26.33	0.844	0.252	52.28	0.611	0.393	1.436 (12)
FOV-KPN [ICCV21] [ 4 ]	4.01	19.00	9.68	26.34	0.824	0.184	50.27	0.631	0.422	1.502 (6)

Table 2: Quantitative performance of methods across

5

aberration severity levels. O. P. denotes overall performance. The ranking in parentheses represents the overall performance ranking of the methods. The number of parameters, FLOPs, and inference time of each method are also listed. The image with a resolution of

256{\times}256

is used to calculate FLOPs and inference time. The best and second best performances are in bold and underline .

Figure 7: Ranking trends of different methods across aberration severity levels, with severity increasing from L1 to L5.

Figure 8: Visual comparison of five representative methods at $3$ levels of aberration severity. We provide RMS/ODE values to facilitate a better understanding of the relationship between quantitative values and qualitative outcomes.

Figure 9: Line charts of different methods under five levels of spatial uniformity (

U_{s}

) and five levels of chromatic aberration severity (

U_{c}

Work	Key Specs (FoV, F/#, Elem.)	Test Set (# Lenses)	Lens Type	Aberration Quantification
Li et al. [ 24 ]	Unspecified	Small / Unknown	Unspecified	N/A (Random)
Gong et al. [ 14 ]	Unspecified	$\sim$ 5	Unspecified	N/A (Random)
OmniLens [ 20 ]	Discrete (6 Specific Sets)	6 (Handcrafted)	Spherical + Aspherical	Spot RMS Radius
UniCAC (Ours)	40 $\sim$ 80 ∘ , F/2 $\sim$ 4, 1 $\sim$ 6 Elem.	120 (Unseen)	Spherical + Aspherical	ODE (Proposed)

Table 3: Comparison with existing CAC benchmarks. UniCAC features strictly defined optical design ranges and a large-scale stratified test set, whereas prior works rely on unspecified or discrete lens samples.

Figure 10: (a) Examples of automatically designed lenses with diverse specifications. (b) t-SNE visualization of aberration distributions in our benchmark. Best viewed when zoomed in.

Figure 11: More comparison with Zemax . For additional validation, we test the simulation system using Cooke triplet lens and Double-Gauss lens.

Figure 12: Comparison between the real aberration image and the extracted pseudo-GT image.

Figure 13: Pipeline for OIQE calculation.

Figure 14: Overall performance comparison of different methods across five aberration severity levels. To enhance visual clarity, we normalize PSNR to a range of

20dB

30dB

and transform LPIPS values by subtracting them from 1.

Figure 15: Overall performance comparison of different methods under

5

image spatial uniformity levels.

Figure 16: Overall performance comparison of different methods under

5

image chromatic aberration levels.

Method	GAN	Codebook	PSNR $\uparrow$	LPIPS $\downarrow$	ClipIQA $\uparrow$
FeMaSR	✗	✗	26.83	0.147	0.491
	✓	✗	26.86	0.145	0.512
	✓	✓	26.94	0.136	0.520

Table 4: Ablation results of training scheme and clean image prior

Method	FoV	PSNR $\uparrow$	LPIPS $\downarrow$	ClipIQA $\uparrow$
DFUnet	✗	25.88	0.253	0.377
DFUnet	✓	26.33	0.252	0.393
Method	PSF	PSNR $\uparrow$	LPIPS $\downarrow$	ClipIQA $\uparrow$
PART	✗	27.22	0.264	0.370
PART	✓	28.10	0.228	0.389

Table 5: Ablation results of optical prior

CAC Dataset	Model	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	OIQE $\uparrow$	ClipIQA $\uparrow$
AODLib [ 20 ]	FOV-KPN [ICCV21] [ 4 ]	23.90	0.756	0.244	74.51	0.477	0.430
	SwinIR [ICCVW21] [ 26 ]	25.05	0.793	0.226	65.18	0.418	0.418
	FeMaSR [MM22] [ 2 ]	24.15	0.748	0.235	67.73	0.568	0.476
UniCAC (Ours)	FOV-KPN [ICCV21] [ 4 ]	26.34	0.824	0.184	50.27	0.631	0.463
	SwinIR [ICCVW21] [ 26 ]	27.22	0.848	0.264	54.41	0.513	0.370
	FeMaSR [MM22] [ 2 ]	26.94	0.841	0.136	34.59	0.722	0.520

Table 6: Quantitative evaluation of the UniCAC .

Metric	Aberration Severity	Model
Metric	Aberration Severity	PART	DFUnet	FOV-KPN	MIMOUnet	NAFNet	Restomer	Uformer	FeMaSR	PromptIR	DiffBIR
PSNR $\uparrow$	L1	29.22	27.57	28.02	28.24	28.47	27.77	28.87	28.05	27.77	28.14
	L2	28.82	27.07	27.03	28.22	28.65	27.77	28.66	27.53	27.63	28.08
	L3	28.29	26.45	26.33	27.52	27.98	27.33	28.10	26.91	27.35	27.84
	L4	27.99	25.81	26.12	27.39	27.53	27.04	27.98	26.68	26.94	27.77
	L5	26.19	24.75	24.20	25.43	26.25	25.29	26.12	25.53	25.12	26.42

Table 7: Quantitative results under

5

different aberration severity levels.

Metric	Spatial Uniformity	Model
Metric	Spatial Uniformity	PART	DFUnet	FOV-KPN	MIMOUnet	NAFNet	Restomer	Uformer	FeMaSR	PromptIR	DiffBIR
PSNR $\uparrow$	Us1	29.22	26.87	27.10	28.30	28.66	27.98	28.89	27.56	27.78	28.18
	Us2	27.89	26.10	26.44	27.56	27.95	27.22	28.17	26.98	27.07	27.89
	Us3	27.73	25.97	26.17	27.07	27.50	26.76	27.74	26.61	26.69	27.41
	Us4	27.37	25.61	25.67	26.71	26.92	26.62	27.42	26.44	26.64	27.37
	Us5	26.43	24.75	25.19	25.65	26.06	25.42	26.11	25.49	25.35	26.51

Table 8: Quantitative results under

5

different spatial uniformity levels.

Computational Imaging

Computational imaging combines optical and sensor design with algorithms to recover or infer scene information from indirect, compressed, or noisy measurements.

See all 123 papers

LensBridge: Frequency-Guided Compound Degradation Adaptation for Lens Aberration Correction and Veiling Glare Removal
Sep 29, 2026Xiaolong Qian, Zhonghua Yi, Qi Jiang +4Image Restoration Diffusion Models for Image Restoration
Zhejiang University · Hunan University
AbsoluteDegradation: A Physics-Inspired Synthetic Film-Degradation Pipeline and Archival Film Restoration Benchmark
Jul 2, 2026Mikołaj Jastrzębski, Dawid Glinkowski, Dawid Zieliński +3Synthetic Data Augmentation Video Restoration
Wrocław University of Science and Technology
UniDAC: Universal Metric Depth Estimation for Any Camera
Mar 28, 2026Girish Chandar Ganesan, Yuliang Guo, Liu Ren +1Metric Depth Estimation Depth Estimation
Michigan State University · Bosch Research North America & Bosch Center for AI (BCAI) · University of North Carolina at Chapel Hill

Sep 29, 2026J/K move · Enter open · S save

LensBridge: Frequency-Guided Compound Degradation Adaptation for Lens Aberration Correction and Veiling Glare Removal

Xiaolong Qian, Zhonghua Yi, Qi Jiang, Kailun Yang+3

Zhejiang University · Hunan University

Image Restoration Diffusion Models for Image Restoration Physics-Guided Image Restoration

Open paper AskPDF arXiv

Simplified optical systems often exhibit residual lens aberrations and Veiling Glare (VG), resulting in spatially varying blur and contrast reduction. Large-scale Lens Libraries (LensLib) enable reusable aberration correction models by covering diverse Point Spread Functions (PSFs), but their aberration-only training distribution does not include target-specific veiling glare. Extending such foundations to compound degradation is challenging because realistic target-system compound pairs are difficult to obtain. To address this challenge, we propose LensBridge, a two-stage framework that first establishes a reusable aberration correction foundation and then adapts it to compound optical degradation using only a few unpaired target observations. In Stage I, we build a PSF-aware one-step diffusion foundation by constructing discrete degradation priors from LensLib PSFs and learning to retrieve them directly from aberrated images, enabling PSF-aware correction without requiring explicit PSF at inference. In Stage II, we adapt this foundation to compound degradation through frequency-domain guidance. At the data level, Frequency-guided Degradation Completion (FDC) transfers target low-frequency characteristics to LensLib aberrated images while preserving aberration structures to synthesize compound training pairs; at the model level, Frequency-guided Pseudo Decomposition (FPD) forms aberration- and VG-dominant pseudo observations to condition separate adaptation branches. Extensive experiments across multiple optical systems demonstrate that LensBridge effectively extends reusable aberration correction foundations to joint aberration correction and veiling glare removal without target-system paired supervision. All code will be available at https://github.com/XiaolongQian/LensBridge.

Abstract

Computational Imaging

Explore similar work

LensBridge: Frequency-Guided Compound Degradation Adaptation for Lens Aberration Correction and Veiling Glare Removal

AbsoluteDegradation: A Physics-Inspired Synthetic Film-Degradation Pipeline and Archival Film Restoration Benchmark

UniDAC: Universal Metric Depth Estimation for Any Camera