Introduction

Metalenses are planar metasurfaces composed of dense arrays of subwavelength scatterers that tailor the wavefront by imparting spatially varying phase, amplitude, and polarization 1,2,3. In contrast, as illustrated in Fig. 1a, conventional refractive singlets and compound lens stacks rely on millimeter–centimeter-scale thickness and multiple glass elements to achieve a given numerical aperture and to correct aberrations 3,4. At high numerical apertures (NA), these refractive objectives often become bulky and inefficient 5. They require thick, heavy glass stacks and carefully optimized multi-element modules to keep images sharp and to avoid characteristic forms of blur and distortion 6,7.

By implementing a customized wavefront within a subwavelength-scale thickness, metalenses offer the potential to replace or complement such bulky optics with compact, wafer-scale elements that support high-NA and wide field of view 8,9,10. Successful demonstrations include broadband or achromatic metalenses, multifunctional metasurface cameras, and compact modules for microscopy and wearable or mobile imaging 11,12,13,14,15,16,17,18.

Instead of bending light through bulk refraction in thick glass elements, metalenses implement the required phase profile by designing meta-atoms to satisfy a desired wavefront. This meta-atom–based phase control, realized through subwavelength scatterers, enables high NA in a thin, lithography-compatible form factor and relaxes packaging constraints compared with multi-element refractive stacks 19,20. However, individual meta-atoms exhibit strongly wavelength-, incidence-angle-, and polarization-dependent responses 21,22,23. In particular, the point-spread function (PSF), which captures the impulse response of the lens, can vary significantly with wavelength, field angle, and polarization, leading to pronounced chromatic dispersion and rapidly changing image quality across the field of view 24,25. As a result, metalenses are sensitive to fabrication tolerances, alignment errors, and deviations from the design spectrum 26. For general-purpose imaging, these effects manifest as design-specific aberrations, wavelength-dependent focal shifts, field-dependent blur, edge darkening (vignetting), and spatially varying distortion that degrade image quality and hinder feature detection, recognition, and tracking 27.

Despite impressive progress in metalens design, several challenges remain before metalenses can serve as drop-in replacements for conventional camera modules. On the optics side, dispersion-engineered and achromatic metalenses can partially correct chromatic focal shifts 11,13. However, they typically trade off bandwidth 21,28, NA 23, and field of view 6,29, or require complex multi-layer 23 or multi-surface architectures that are difficult to fabricate at scale 30. Hybrid systems that combine metalenses with refractive corrector elements can improve aberration performance, but they sacrifice some of the thickness and weight advantages that motivate metasurface optics in the first place 31. Moreover, metalens performance is often tightly coupled to a specific wavelength band, sensor, and packaging configuration, which limits portability across applications and product generations 32.

To address these limitations, there has been growing interest in computational metalens imaging that jointly considers optics and post-processing 33,34,35,36. As illustrated in Fig. 1b, physics-based pipelines estimate or measure the PSF of a given metalens and then apply deconvolution or model-based restoration to reduce blur and color fringing 24,37. However, for metalenses intended for commercial, high-volume production, software correction of metalens-specific aberrations still requires large image datasets captured with the manufactured metalens modules 24. Building such per-device datasets across wafers, production batches, and alignment conditions is a labor-intensive, time-consuming process that is difficult to integrate directly into standard high-throughput manufacturing flows.

Fig. 1
Fig. 1
Full size image

Conventional, metalens, and proposed augmentation pipelines. (a) Conventional imaging system: A compound refractive module with multiple lens elements produces high-quality images at the cost of increased axial length and mass. (b) Ultra-compact metalens imaging system: A single metalens yields raw captures with chromatic fringes, field-dependent blur, and spatial distortion; a learned restoration network is required to obtain a conventional-looking image. (c) Proposed augmentation pipeline: A PSF-free, device-conditioned image-to-image translator converts ordinary photographs into metalens-style renders that reproduce chromatic and field-dependent artifacts. The resulting synthetic pairs enable training of restoration models without additional metalens data collection, providing many restored images for downstream analysis.

In this work, we target this data bottleneck from a complementary direction. Instead of repeatedly collecting massive paired datasets for every new metalens design, we develop a few-shot data-augmentation framework that learns to synthesize metalens-style images, defined as ordinary photographs transformed to mimic the characteristic aberrations and degradations of a given metalens. Given a clean input image captured by a conventional lens, the proposed image-to-image translation model generates a plausible counterpart that exhibits the chromatic aberration, field-dependent blur, and spatial distortion produced by a target metalens. Trained on a modest calibration set of metalens/conventional pairs, the model then scales to produce large volumes of synthetic training data without explicit PSF measurement, wave-propagation simulation, or exhaustive metalens image acquisition (Fig. 1c). In this way, our approach combines the advantages of deep learning-based image restoration with the compact, wafer-thin form factor of metalenses, making software correction of metalens aberrations more compatible with high-volume manufacturing.

Results

We now evaluate how well the proposed image-to-image translator can synthesize realistic metalens-style degradations and how useful the resulting synthetic images are for metalens image restoration. Here, we use metalens-style to denote ordinary photographs transformed to mimic the characteristic chromatic fringes, field-dependent blur, and spatial distortions produced by a target metalens.

From a computer-vision perspective, our task lies between data augmentation and image-to-image translation. Classical augmentation pipelines (random crops, flips, rotations, color jitter, and policy-based schemes) are designed to increase the stochastic diversity of training sets 38,39, while restoration and super-resolution networks are trained to map low-quality inputs to high-quality outputs 40,41. In meta-optical imaging, recent deep-learning-based systems have mainly followed this enhancement direction, restoring strongly aberrated metalens captures to near-conventional-lens quality using hundreds to thousands of paired images per device 24,26. In contrast to these enhancement models, we focus on the reverse direction, mapping clean images to degraded metalens-style images. We deliberately degrade conventional images to resemble the outputs of a specific, mass-producible metalens, aiming to generate large synthetic datasets that can replace or significantly reduce the need for labor-intensive paired collections.

We organize our evaluation into four parts. First, we quantify the quality of the proposed metalens-style synthesis on a test set of paired conventional and metalens images using standardized image metrics. Second, we benchmark the translator against conventional restoration and super-resolution baselines trained under identical preprocessing and evaluation protocols to verify that a degradation-focused model is necessary to reproduce metalens-specific artifacts. Throughout this section, the restoration/super-resolution baselines are trained and evaluated in the clean \(\rightarrow\) metalens direction, and their performance is interpreted in terms of degradation-synthesis fidelity to the ground-truth metalens images rather than enhancement quality. Third, we perform ablation studies on the loss formulation and upsampling strategy to isolate their effects on chromatic aberration reproduction, field-dependent blur, and texture retention. Finally, we assess the utility of the synthesized datasets for metalens image restoration by training a standard restorer exclusively on translated images and measuring its generalization to real metalens captures. This last experiment directly tests our central hypothesis that physically informed, metalens-conditioned augmentation can alleviate the data bottleneck in deep-learning-driven meta-optical imaging.

Fig. 2
Fig. 2
Full size image

Data synthesis and restoration workflow for metalens imaging. (a) Pix2Pix-based conditional translation framework. A U-Net generator learns to convert conventional-lens captures into synthesized metalens-style images that exhibit realistic chromatic and field-dependent aberrations. A pair of discriminators evaluates realism at multiple patch scales, with the generator and discriminators updated alternately for stable training. (b) Restoration pipeline using synthetic data. The restoration network is trained exclusively on paired data composed of synthesized metalens-style images and their corresponding conventional-lens references, producing restored outputs in the conventional-lens domain without real metalens supervision.

Image synthesis for metalens imaging

We first evaluate the fidelity of the proposed metalens-style synthesis on a set of paired conventional and metalens-captured images. As detailed in Fig. 2, our translator is trained on a small calibration set of paired captures and then applied to unseen conventional photographs to generate metalens-style renderings. As illustrated in Fig. 2, the generator employs a U-Net architecture 42, leveraging skip connections to effectively capture detailed information between the input and output images, thereby generating accurately transformed images. While the original Pix2Pix 43 model typically employs an L1 loss function (mean absolute error, MAE) 44, this study utilizes an L2 loss function (mean squared error, MSE) 45 to preserve realistic image textures better. The Discriminator employs a robust 1x1 PatchGAN architecture 43, which enhances its ability to discern detailed textures and assess the realism of generated images with respect to colorfulness.

Table 1 Acquisition time for real metalens datasets versus synthetic metalens-style datasets for \(N = 600\) images. The synthetic generation time is assumed to be \(\sim 50\) ms per image on a single NVIDIA RTX 4090 GPU.
Full size table

Once trained, the translator can convert arbitrary conventional photographs into metalens-style images that exhibit realistic optical aberrations, including chromatic fringes and field-dependent blur. In our implementation, generating a dataset of 600 synthetic metalens-style images takes on the order of 30 s, compared with roughly 30 min for real metalens acquisition (Table 1).

Fig. 3
Fig. 3
Full size image

Comparison of metalens-style synthesis across models. The proposed Pix2Pix-based translator most faithfully reproduces chromatic aberration and field-dependent blur without introducing noticeable perceptual artifacts, yielding more vivid color and realistic contrast than restoration-based or super-resolution methods, which often produce desaturated and over-smoothed regions.

To quantify the fidelity of our synthetic metalens augmentation, we compare three classes of methods–image-to-image translation, image restoration, and super-resolution–on a test dataset of paired conventional and metalens-captured images. Each model is evaluated using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS). As shown in Table 2, Pix2Pix outperforms all competitors, achieving a mean PSNR of 29.37 dB, a mean SSIM of 0.9877, an LPIPS score of 0.0477, and an LPIPS-VGG score of 0.1168. Qualitative results in Fig. 3 corroborate the quantitative findings: Pix2Pix faithfully reproduces chromatic fringes, field-dependent blur, and spatially varying distortions, closely matching ground-truth metalens captures without introducing perceptual artifacts. In comparison, restoration-based methods tend to over-sharpen or smooth critical features, and super-resolution models like LAPAR 46 are unable to emulate realistic aberrations.

Table 2 Performance comparison of image models on metalens data, grouped by task. For each metric, the first value denotes the mean and the second value denotes the standard deviation; higher SSIM and PSNR are better, whereas lower LPIPS is better.
Full size table

Restoration using synthetic training data

We next investigate whether realistic restoration of metalens photographs can be achieved using only a limited amount of real metalens data together with the proposed augmentation. We first train the translator on about 600 real metalens images. Then, images from the Real-SR dataset 52 are cropped to a resolution of \(512\times 360\) and converted into metalens-style images, yielding 500 synthetic samples. The restoration network is trained on these synthetic data, along with 1,100 real metalens images. Training pairs consist of either a real metalens image or a metalens-style synthetic image paired with its corresponding clean reference image. Restoration performance is evaluated on real metalens photographs that are not used during training.

Evaluation on real metalens images

Generalization was assessed on a collection of photographs captured with the metalens that were not used in translator training or restoration training. Although the restorer is trained only on synthetic pairs generated from a limited amount of real measurements, the augmentation-only restorer consistently recovers the main structures and overall color balance in real metalens images. As shown in the visual comparison in Fig. 4, the Pix2Pix-based model produces more natural, structurally faithful outputs than the other baselines, with better fine-detail preservation and perceptual quality (SSIM, LPIPS). While reconstructions tend to be somewhat softer than those obtained by a restorer trained directly on real pairs, the results demonstrate that synthetic augmentation alone can still provide meaningful restoration for real metalens inputs.

Fig. 4
Fig. 4
Full size image

Qualitative restoration results on real metalens photos using a model trained on real and augmented data. Right: restoration results for real inputs. Major structures and color tones are recovered even with only limited real-pair supervision; finer textures remain challenging, especially near the image edges.

Quantitative results

For quantitative validation, we evaluated restoration models on 70 real metalens test pairs using PSNR, SSIM, and LPIPS metrics. All restorers were trained exclusively on synthetic pairs composed of translated metalens-style inputs and their corresponding conventional-lens references; no real metalens/conventional pairs were used during training.

Table 3 Restoration performance on the real metalens + synthetic metalens test set. Higher values indicate better performance for PSNR and SSIM, while lower values are better for LPIPS.
Full size table

As summarized in Table 3, the restorer achieves a PSNR of 19.159, SSIM of 0.745, and LPIPS of 0.452. This outperforms strong restoration baselines trained under the same synthetic-only regimen: the next-best PSNR (17.640 dB) and LPIPS (0.543) are obtained by Restormer 41, while the next-best SSIM (0.732) is achieved by Palette 47. In other words, our model improves PSNR by roughly 1.5 dB over the strongest baseline and reduces LPIPS by about 0.09 in absolute terms, indicating better perceptual agreement with conventional-lens references even though it is trained solely on synthetic metalens-style pairs without access to any real metalens/conventional pairs.

These quantitative gains are consistent with the qualitative results in Fig. 4: the our Pix2Pix-based restorer recovers global structure and color tone more faithfully than competing methods and suppresses many of the severe chromatic aberration and field-dependent blur present in the raw metalens captures. Taken together, these results support the use of synthesized metalens-style data as effective training targets for restoration in metalens imaging.

Observed failure modes and limitations

A consistent artifact appears near the image periphery. At large field angles, edges and fine detail can be over-blurred in the restored outputs. This behavior is attributed to underrepresented position-dependent blur in the augmented training inputs. Because the current augmentation is not explicitly conditioned on spatial coordinates, the restorer tends to learn a spatially uniform correction that does not capture edge-of-field behavior 24. In practice, boundaries in the synthetic images are not blurred as strongly or as anisotropically as those observed in real captures, and the restoration model reproduces this discrepancy. Remedies include injecting normalized spatial coordinates into both translator and restorer, sampling training patches by field angle, and employing spatially varying losses. Mixing a small number of real pairs during restorer training is also expected to reduce any residual gap between synthetic and real data while keeping the collection cost low.

Fig. 5
Fig. 5
Full size image

Effect of loss and upsampling on metalens-style synthesis. We compare L1 versus L2 objectives and transpose-convolution versus bilinear upsampling; the L2 + bilinear variant best suppresses checkerboard artifacts while preserving subtle aberration features.

Ablation study

Effect of loss and upsampling

We isolate the effects of loss formulation and upsampling approach through an ablation study illustrated in Fig. 5. Specifically, we train the Pix2Pix 43 model with either an L1 or L2 objective and replace transposed convolutions with bilinear upsampling. Quantitative and qualitative assessments reveal that the L2 objective paired with bilinear interpolation reduces checkerboard artifacts and preserves subtle aberration features more faithfully than other combinations. In contrast, models trained with an L1 objective or using transpose convolution introduce visible grid patterns or over-smooth fine details, underscoring the importance of loss choice and sampling method for realistic metalens augmentation.

Table 4 Effect of training set size on metalens-style synthesis performance. Results are reported for PSNR, SSIM, and LPIPS on the test set; higher PSNR and SSIM and lower LPIPS indicate better performance.
Full size table

Effect of training set size

We examined data efficiency by training the translator with 100, 300, 500, and 600 paired samples, as summarized in Table 4. As the set size increases, PSNR and SSIM improve while LPIPS decreases, indicating progressively better agreement with real metalens captures. With smaller training sets (100–300 pairs), the generator largely preserves hue and overall style but struggles near the image edges: straight lines bend or break, thin contours fray, and high-frequency textures collapse into overly smooth bands. As the training pool grows, these artifacts are progressively reduced: boundary warping recedes, line continuity improves, and small features such as wires and edge highlights are retained more reliably. At the largest set size (600 pairs), the synthesized images exhibit more stable geometry and sharper micro-details across scenes, indicating that additional data primarily benefits spatial fidelity rather than color rendition and reduces rare failure cases at object edges and near field boundaries.

Discussion

We formulate deterministic aberration synthesis for metalens imaging as an image-to-image translation problem. The learned translator maps clean images to metalens-style renderings that reproduce chromatic fringing, field-dependent blur, and spatial distortions, and the resulting synthetic data enable training and evaluation of restoration models on real captures. Rather than enumerating additional metrics, this section interprets the implications, situates the method relative to prior work, and outlines directions for improvement. With respect to prior work in image-to-image translation, most methods emphasize stochastic diversity and multi-modality 53,54,55. By contrast, our objective is metalens-conditioned fidelity, that is, faithfully reproducing the aberrations of a specific metalens from a clean input image.

Generality across metalens designs and operating conditions. In this study, we train and evaluate the translator using data captured with a single metalens design (consistent with the lens used in our prior metalens image restoration setting) to maintain a controlled and directly comparable experimental setup. Because metalens imaging characteristics can vary across designs, fabrication tolerances, and system-level alignment, the current validation does not establish zero-shot generalization to arbitrary metalenses. In practical deployment, we expect a modest re-calibration step (e.g., fine-tuning with a small paired conventional/metalens calibration set) when the metalens design or fabrication/alignment regime changes. As follow-up work, we aim to extend the evaluation to multiple metalens designs and diverse conditions to systematically assess robustness and generality.

The proposed formulation reduces or even eliminates the need for dedicated physical calibration experiments, while remaining compatible with physics-based pipelines. While classical PSF modeling or rigorous simulation provides interpretability, such approaches can be costly or brittle for metasurfaces whose PSFs vary rapidly with field angle and wavelength and depend on nanoscale geometry.

Our approach offers three practical benefits. First, it lowers the entry cost for new devices by amortizing limited calibration into a generator that synthesizes large labeled sets without explicit PSF estimation. Second, it preserves subtle, spatially varying artifacts that generic style-transfer objectives tend to suppress, thereby improving the ecological validity of downstream training. Third, it provides a practical framework for optics–algorithm co-design: after a modest per-device calibration, scene-matched synthetic datasets enable controlled, low-cost iteration over metalens form factors and post-capture processing.

Taken together, these steps aim to sharpen peripheral detail, stabilize color under severe fringing, and position PSF-free aberration synthesis as a practical complement to physics-based modeling for metalens imaging. Beyond the laboratory, such a framework could help compress multi-element refractive modules into single-metalens front-ends with learned digital correction, reducing lens count, thickness, and assembly complexity in mobile and embedded cameras 56. By providing metalens-specific synthetic datasets on demand, it can shorten the iteration loop between optical design 57,58, wafer-level manufacturing 59, and on-device imaging 7, thereby supporting co-designed hardware–software imaging systems in which flat optics and neural networks jointly deliver the performance traditionally achieved by bulky glass stacks.

Methods

Evaluation protocol

To evaluate the proposed augmentation framework, we compare the images generated by our Pix2Pix 43-based translator with outputs from an image-to-image translation baseline 47, two state-of-the-art restoration models 41,49, and leading super-resolution models 46. Although restoration and super-resolution networks are originally designed for enhancement (low-quality \(\rightarrow\) high-quality), we include them here as enhancement-oriented controls and retrain/use them in the clean \(\rightarrow\) metalens synthesis direction under identical paired-data supervision and preprocessing. This benchmark therefore evaluates how well each model can reproduce metalens-specific degradations; accordingly, the metrics in Table 2 quantify synthesis fidelity to the ground-truth metalens captures rather than enhancement quality.

For quantitative assessment, we compute PSNR, SSIM, and LPIPS between each model’s synthetic outputs and the corresponding ground-truth metalens captures. Table 2 reports these metrics on a test set, and Fig. 3 presents representative qualitative examples. To verify that our synthetic images serve as valid training data, we train a standard restoration network using only the generated dataset and then evaluate its performance on real metalens images. We measure PSNR, SSIM, and LPIPS for these reconstructions and analyze their error distributions. The resulting scores demonstrate that models trained on our augmented data achieve competitive reconstruction quality, confirming that the generated images are effective for downstream learning in metalens imaging tasks.

Dataset detail

In our framework, we use two types of image data: (i) ground-truth images captured with a conventional lens, and (ii) their corresponding captures obtained with a metalens. We employ the same dataset as in our previous work by Seo et al., “Deep-learning-driven end-to-end metalens imaging” 24, which comprises 600 training pairs and 70 test pairs.

Hyperparameters

Training hyperparameters followed settings in the Table 5.

Table 5 Training setup for the Pix2Pix-based metalens-style translator.
Full size table