betaFeedback

WorksComputer Science

STAR++: A Physical-Aware Benchmark and Framework for Astronomical Super-Resolution across Multiple Filters

Abstract

Astronomical super-resolution (ASR) aims to reconstruct high-resolution star fields from low-resolution telescope observations across multiple filters for astrophysical analysis. However, developing a physics-aware ASR framework faces two challenges: 1) Data view: no large-scale benchmark supports field-level ASR, especially on unseen filters. 2) Method view: Existing ASR methods often fail to preserve stellar flux across spatial resolutions. Moreover, ASR requires reconstructing high-frequency components, including accurate star locations and profiles, making it more ill-posed than natural-image SR. To tackle these problems, we introduce STAR++, a large-scale benchmark comprising approximately 1.7 million flux-consistent image pairs across 36 filters, derived from Hubble Space Telescope (HST) observations. STAR++ uses a flux-preserving data generation pipeline to ensure photometric consistency across resolutions and provides the Flux Error (FE) metric to evaluate ASR models from a physical perspective. To address method-level challenges, we propose a Flux-Invariant Super-Resolution Plus (FISR++) model that generalizes to unseen filters while accurately reconstructing flux-consistent high-resolution images. Beyond the flux-preserving modules, FISR++ introduces two novel components. Specifically, a Filter-Aware Adapter (FAA) incorporates filter metadata to adapt image representations to varying spectral and optical conditions. In addition, a Spatial Ambiguity Disentangle Loss (SADL) models spatial ambiguity as pixel-shift uncertainty, mitigating the smoothing of learned conditional mean reconstructions caused by this ambiguity. Extensive experiments demonstrate that FISR++ consistently outperforms existing ASR methods in reconstruction quality and flux accuracy, while generalizing to unseen filters, highlighting its effectiveness for photometrically faithful ASR.

  • Guohang Zhuang and Jiaze Li contributed equally to this work.

This work was supported by the JC STEM Lab of AI for Science and Engineering, funded by The Hong Kong Jockey Club Charities Trust, the Research Grants Council of Hong Kong (Project No. CUHK14213224).

Introduction

Higher resolution is critical to astronomical observation, as it reveals finer astrophysical structures and supports more precise scientific measurements [1, 2, 3]. Achieving higher spatial resolution traditionally relies on advances in telescopes and observing systems, often at substantial instrumental and operational cost [4, 5]. Super-resolution (SR) offers a complementary computational approach to reconstruct high-resolution images from lower-resolution observations [6, 7, 8], and has shown promise in astronomical imaging [9, 10]. However, astronomical images contain physically meaningful information, such as the flux of celestial sources. Therefore, developing a physics-aware ASR framework is essential for accurately reconstructing star fields while preserving their physical fidelity.

Physics-aware ASR requires cross-filter generalization, flux consistency (i.e., consistent stellar brightness across spatial resolutions), and accurate reconstruction of stellar structures. However, existing ASR datasets and methods still face limitations from two perspectives to achieve the aforementioned goals. 1). Data perspective: Existing ASR benchmarks cover a limited number of filters and lack systematic evaluation on unseen filters [11, 12, 13], as shown in Tab. I. Moreover, existing ASR networks primarily rely on image inputs without explicitly incorporating filter-specific physical properties such as wavelength and bandwidth, limiting their adaptability to different observation conditions. 2). Method perspective: It mainly includes two challenges of flux inconsistency and spatial ill-posedness. (i) Flux inconsistency: In the real world, telescopes under different observation resolutions have a flux consistency relation [14, 15]. Specifically, although a celestial object has different levels of distortion under low resolutions, it almost has the same total flux as in high resolutions because of the telescope imaging principle [14, 16]. However, existing datasets have significant drifts from this property because they directly use simple interpolation [17, 18], suitable for natural images but conflicts with astronomical observation. (ii) Spatial ill-posedness: Unlike natural image super-resolution, which can use surrounding context to restore large objects, ASR must recover isolated, tiny point sources like stars. Without contextual clues, a single low-resolution pixel can correspond to many possible high-resolution stellar positions and profiles, making the recovery severely ill-posed. To minimize error under this spatial ambiguity, common likelihood training simply averages all candidate positions, which flattens the peak intensity, leading to over-smoothed restoration and inaccurate source flux.

To support comprehensive, large-scale, multi-filter ASR research across diverse astronomical fields, we introduce STAR++, a flux-consistent benchmark for field-level astronomical super-resolution. It extends STAR [19] from one filter to 36 filters spanning the 222.4–906.0 nm band. STAR++ contains nearly 1.7 million flux-consistent LR-HR pairs, generated from 2,019 Hubble Space Telescope (HST) [16] science images. The source images cover diverse astronomical fields, covering a large range of star fields and on average containing 55 celestial objects per image and complex scenarios including multiple celestial objects, cross-object interactions, and weak lensing phenomena, as illustrated in Figure 1. STAR++ retains the flux-consistent data generation pipeline of STAR to preserve the photometric properties of astronomical observations across resolutions. Further, we provide Flux Error (FE) to evaluate reconstruction quality from a physical perspective.

Comparison of previous datasets and STAR++ dataset examples, showing individual galaxies, cross-object interactions, weak lensing phenomena, dark matter halos, galaxies, spiral galaxies, and dense star fields

Figure 1. Comparison of previous datasets and ours, highlighting the richer astronomical structures and phenomena captured by STAR++, such as cross-object interactions, weak lensing, Dense star fields, and spiral galaxies. The STAR++ examples are visualized as RGB composites, with red, green, and blue filter labels indicating the filters mapped to the corresponding color channels.

Building on STAR++, we propose a novel field-level ASR model, Flux-Invariant Super-Resolution Plus (FISR++), to improve cross-filter generalization and reconstruction accuracy while maintaining flux consistency. In addition to the flux-preserving Flux Guidance Generation (FGG) and Flux Guidance Controller (FGC) modules and the Flux Consistency Loss (FCL) [19], FISR++ incorporates two novel modules, namely a Filter-Aware Adapter (FAA) and Spatial Ambiguity Disentangle Loss (SADL). Specifically, the FAA incorporates filter metadata, allowing the model to recognize filter and adapt filter-specific knowledge to restore high resolution image across different filters. This filter-aware design enables the model to adapt its super-resolution patterns according to the filter condition, thereby improving its generalization to unseen filters. In addition, we propose a Spatial Ambiguity Disentanglement Loss (SADL) to accurately restore stellar positions and profiles. During training, SADL explicitly models spatial ambiguity as pixel-shift uncertainty and separates it from image reconstruction. Learning this uncertainty independently keeps the model from averaging across candidate positions, so it can pinpoint exact stellar peaks and preserve sharp profiles. At inference, we simply drop the uncertainty term to produce clean, sharp high-resolution images.

In summary, the key contributions of this paper are as follows:

  1. STAR++ Benchmark: We present STAR++, a multi-filter, flux-consistent ASR benchmark comprising approximately 1.7 million LR-HR pairs with 36 HST filters.

  1. Flux-Invariant Super Resolution Plus (FISR++) Model: We propose FISR++, extending FISR with a Filter-Aware Adapter (FAA) that incorporates physical filter priors and a Spatial Ambiguity Disentanglement Loss that accounts for

DatasetSizeType# FiltersDownsamplingField-levelFlux ConsistencyUnseen-filter Eval.
AstroSR [11]2,000Galaxy5××××
Shan et al. [12]9,383Galaxy1××××
Song et al. [17]1,597Solar1✓×××
DiffLense [18]2,880Galaxy1××××
Luo et al. [13]17,000Galaxy3××××
Li et al. [20]14,604Galaxy3✓×××
STAR54,738Star field1✓✓✓×
STAR++1,697,816Star field36✓✓✓✓

Table 1. Comparison of existing ASR datasets.

pixel-wise uncertainty and sub-pixel misalignment.

(3) Extensive experiments demonstrate that FISR++ achieves state-of-the-art performance on STAR++, with a PSNR of 35.42 dB and a Flux Error (FE) of 0.4341 on unseen filters. Compared with the conference-version FISR, FISR++ improves PSNR by 1.1 dB and reduces FE by 0.0963.

Difference from the conference paper: The preliminary version of this manuscript is conference paper STAR [19]. Compared with the conference version, the improvements of the STAR++ and FISR++ are listed below: 1) We expand STAR from a single filter to 36 HST filters, increasing the dataset from 54,738 to approximately 1.7 million flux-consistent LR–HR pairs. 2) The proposed Filter-Aware Adapter (FAA) incorporate filter metadata into image representations. We also adopt an asinh intensity representation to compress the dynamic range and better balance the contributions of bright and faint sources during training. 3) We introduce a Spatial Ambiguity Disentangle Loss that explicitly models pixel shifts and pixel-wise reconstruction uncertainty to improve the positional accuracy of reconstructed high-frequency structures and reduce spatial ambiguity. 4) We conduct additional experiments to evaluate generalization to unseen filters and compare the performance of models trained on STAR and STAR++. Furthermore, FISR++ achieves a PSNR of 35.42 dB and a Flux Error (FE) of 0.4341 on STAR++, representing a 1.1 dB improvement in PSNR and a 0.0963 reduction in FE over FISR, thereby demonstrating state-of-the-art performance.

Related Works

Astronomical Super-Resolution Benchmarks

Astronomical super-resolution (ASR) aims to reconstruct high-resolution astronomical observations from low-resolution images. Developing reliable ASR models requires large-scale training data, which has motivated the recent development of several ASR datasets and benchmarks. AstroSR [11] constructed 2,000 paired galaxy images from overlapping SDSS and HSC observations for cross-survey galaxy super-resolution. Shan et al. [12] further collected 9,383 galaxy image pairs for diffusion-based reconstruction, while Song et al. [17] and DiffLense [18] extended ASR to solar images and gravitational-lensing observations, respectively. Luo et al. [13] and Li et al. [20] explored cross-survey ASR with datasets containing 17,000 and 14,604 image pairs, respectively. Compared with these datasets, STAR [19] introduced 54,738 flux-consistent

HST star-field pairs for field-level ASR, covering multiple celestial objects and complex phenomena such as cross-object interactions and weak lensing while preserving physical fidelity. More recently, StelLens [21] constructed a large-scale paired SDSS-HST dataset for learning ground-to-space astronomical reconstruction. FluxFlow [22] introduced a DESI-HST benchmark with 19,500 image pairs and realistic atmospheric degradation. However, despite these advances in dataset scale and astronomical targets, existing ASR benchmarks still cover only a limited number of predefined spectral filters, leaving cross-filter and unseen-filter generalization insufficiently studied.

Super-Resolution Techniques

Super-resolution (SR) techniques aim to reconstruct high-resolution (HR) images from low-resolution (LR) observations. Traditional approaches, including interpolation, deconvolution, and sparse representation [23], recover image details through predefined degradation models or learned statistical relationships. Deep learning has substantially advanced SR through convolutional neural networks (CNNs) [6, 8, 24], while generative adversarial networks (GANs) have improved perceptual reconstruction quality [25, 26, 27]. Transformer-based methods further exploit attention mechanisms to capture long-range dependencies and recover fine image structures [28, 29, 30, 31]. More recently, diffusion-based methods have employed iterative denoising processes to generate high-quality reconstructions [32, 33, 34]. Although these techniques have achieved substantial improvements in image reconstruction, visual similarity and perceptual quality alone do not guarantee the physical fidelity required for astronomical analysis.

Astronomical Super-Resolution Methods

Astronomical image reconstruction has traditionally relied on deconvolution [10, 35], which models the point spread function (PSF) [36], and multi-exposure reconstruction [37, 38], which combines information from multiple observations. Deep learning has subsequently enabled reconstruction of various astronomical targets, including galaxies [39], solar structures [40], and X-ray sources [41]. Beyond individual celestial objects, STAR [19] introduced field-level ASR for complex star fields containing multiple objects with varying brightness. Recent astronomy-specific methods have further incorporated physical information into image reconstruction. @Winvalid STAR [19] introduced flux-guided reconstruction, while Stel-Lens [21] explored observation-physics-driven ground-to-space reconstruction. FluxFlow [22] employed flow matching with observation uncertainty to improve astronomical reconstruction. Nevertheless, robust reconstruction across diverse and previously unseen filters remains challenging, particularly when observation characteristics vary across filters. Moreover, small positional deviations between reconstructed and reference images can introduce substantial pixel-wise errors, limiting the accurate recovery of stellar structures.

Flux Consistency in Astronomical Image Processing

Flux consistency, ensuring that the total light intensity (flux, or photons received per unit area) of celestial objects in processed images matches original observations, is a cornerstone of astronomical image analysis, underpinning reliable photometry and stellar population studies [42], [43]. Historical efforts prioritized flux consistency to preserve measurement accuracy in star clusters and galaxies. The ground-based survey telescope SDSS also follows this principle [44]. However, the complexity of star fields, with diverse brightness and overlapping objects, poses ongoing challenges. Our work advances this field with STAR++, a large-scale star field dataset ensuring flux-consistent image pairs, and novel flux consistency constraints, enhancing the scientific reliability of star field analyses.

STAR++

Following natural SR works, we construct cross-resolution image pairs by downsampling high-resolution images. We choose Hubble Space Telescope (HST)1 survey data as our high-resolution images due to its widely recognized data quality and rich historical data accumulation. Given a high-resolution HST image, we first apply a point spread function (PSF) [36] kernel to simulate the optical blurring effects caused by low-quality telescopes and atmospheric turbulence. Next, we downsample the image by a factor of ss using a flux-conserving scheme. Finally, since both HR and LR images have large spatial dimensions, we divide them into smaller sub-images to facilitate model training. This pipeline generates physically consistent cross-resolution pairs of images, crucial for robust super-resolution model training. The statistics of the resulting dataset are summarized in Table II.

SplitFilter(s)Pivot. λ\lambda (nm)Frac. BWImagesLR–HR pairs
Imagers: WFC3/UVIS (1,106 images); ACS/WFC (913 images)
Sky coordinates: RA 0.07°~359.94°; Dec −89.13°~87.41°
Training (seen)29 filters270.3–843.70.004–0.3521,008922,586
Validation (seen)29 filters270.3–843.70.004–0.352237162,593
Test (unseen-interp.)F435W432.90.0698276,019
F555W533.60.0829347,169
F606W591.20.113379351,326
F625W629.10.0685248,623
Subtotal432.9–629.10.068–0.113606523,137
Test (unseen-extrap.)F218W222.40.056553,500
F225W236.50.0745338,536
F850LP906.00.0576047,464
Subtotal222.4–906.00.056–0.07416889,500
Total36 filters222.4–906.00.004–0.3522,0191,697,816

Table II. Dataset statistics, data splits, and filter characteristics of STAR++, including pivot wavelength and fractional bandwidth. Unseen filters are divided into interpolation and extrapolation subsets according to their pivot wavelengths relative to the training wavelength range.

High-resolution Data Collection

The HST is a space-based observatory designed to capture high-resolution astronomical images across a wide range of wavelengths, from ultraviolet to near-infrared. It provides two kinds of data, including calibration and science. Calibration data is used to correct instrumental effects while the science has a verified quality for scientific research. So we choose scientific data due to its high and reliable quality.

The science data consist of images acquired by various imaging instruments onboard HST. We select two representative high-resolution imagers, including the Advanced Camera for Surveys Wide Field Channel (ACS/WFC) [45] and the Wide Field Camera 3 Ultraviolet-Visible Channel (WFC3/UVIS) [46]. ACS/WFC provides a wide field of view (202′′×202′′202^{\prime\prime} \times202^{\prime\prime}) and high sensitivity over optical wavelengths (370-1100 nm), while WFC3/UVIS provides a field of view (162′′×162′′162^{\prime\prime} \times162^{\prime\prime}) with broader ultraviolet-to-red coverage (200-1000 nm). Their high spatial resolution, large fields of view, and complementary spectral coverage make them well suited for constructing a large-scale multi-filter ASR benchmark.

In practice, HST observations are acquired through multiple filters with distinct wavelength-dependent transmission profiles. Since celestial sources exhibit wavelength-dependent emission, their observed intensities vary across filters, making multi-filter coverage essential for evaluating cross-filter generalization in ASR. STAR++ incorporates 36 HST filters to substantially increase spectral diversity. These filters have pivot wavelengths ranging from 222.4 to 906.0 nm and include wide, medium, narrow, long-pass, and extremely wide filters. As shown in Table II, we reserve seven filters for testing to evaluate generalization, and their corresponding central wavelengths are not represented among the filters used for training. Such diverse filter configurations enable STAR++ to cover multiple filter observation conditions and support systematic analysis of cross-filter generalization.

Figure 2. Schematic diagram of the flux-consistent downsampling process. The workflow illustrates the transformation of HR and LR image pixels into the celestial coordinate system, the computation of overlapping sky regions between HR and LR receptive fields, and the flux calculation for LR pixels using weighted contributions from overlapping HR pixels.

PSF Blurring

Different resolution devices share different PSF blurring. To simulate this phenomenon and increase training data diversity, we adopt two representative PSF models. The first is the Gaussian PSF [47], a simple model widely used to approximate blur caused by atmospheric turbulence or instrumental imperfections in astronomical observations [48, 49]:

PSF⁡Gaussian(x,y)=exp⁡(−x2+y22σ2).(1)\operatorname{PSF}_{\mathrm{Gaussian}}(x,y)=\exp\left(-\frac{x^{2}+y^{2}}{2\sigma^{2}}\right). \tag*{(1)}

The second is the Airy PSF, a device-specific kernel widely used in astrophysics. It captures diffraction effects from a telescope’s circular aperture and describes the instrumental effect that occurs when a telescope resolution changes, making it suitable for space-based instruments like HST [50]:

PSF⁡Airy(r)=[2J1(kr)kr]2.(2)\operatorname{PSF}_{\mathrm{Airy}}(r)=\left[\frac{2J_{1}(kr)}{kr}\right]^{2}. \tag*{(2)}

In these two models, σ\sigma and rr serve as adjustable parameters to control the spread of the blur by modulating the energy dispersion of the filter. In the Gaussian PSF, a larger σ\sigma leads to a less concentrated signal with greater energy spread across the filter, while in the Airy PSF, rr governs the radial extent of energy distribution due to diffraction. We define these parameters based on the telescope’s observed blur characteristics, following Schawinski et al. [39], who used the observed blur to set the PSF parameters for a realistic simulation of hardware-specific degradation effects.

With these two PSFs, we define a blurring setting as a random selection between Airy and Gaussian, each chosen with equal probability, to simulate more complex scenarios.

Flux Consistency Downsampling

After operating PSF blurring, we perform a flux consistency downsampling scheme to obtain realistic low-resolution counterparts, which we detail in the next section.

The flux consistency relation in real-world telescopic observations stems from the telescope imaging principle. The value of each pixel in observational data corresponds to the photon flux captured by its corresponding CCD pixel. Consequently, when imaging the same celestial region, a single CCD pixel in lower-resolution instruments covers a larger spatial area—equivalent to integrating photons from multiple high-resolution CCD pixels. Based on this mechanism, Flux-Consistent Downsampling is implemented in three steps: computing the image-plane coordinates, transferring them to celestial coordinates, and calculating the flux of each LR pixel according to the overlap between HR and LR pixel regions.

Computing image plane coordinates: For a given high-resolution (HR) image with resolution H×WH \times W and downsampling rate ss, we first generate the downsampled low-resolution (LR) grid after valid padding or cropping so that paired HR and LR regions keep an integer scale relation. The resulting LR image size is Hs×Ws\frac{H}{s} \times\frac{W}{s}. These HR and LR pixel coordinates define the image-plane grids used for WCS reprojection.

Transfer pixels to sky: For HR and LR pixel coordinates, we transfer them into the celestial coordinate system as: (u,v)→(ra,dec)(u,v) \rightarrow(ra,dec), where (u,v)(u,v) is a coordinate in the image plane while (ra,dec)(ra,dec) is the longitude and latitude coordinates of the Earth. Note that, each pixel is not an ideal point and actually a rectangle on the image plane. After the mapping, it becomes a quadrilateral surface of the celestial coordinate system. The physical meaning of this quadrilateral surface is the sky area covered by a pixel, denoted as the receptive field here. For the ii-th pixel of the LR image and the jj-th pixel of the HR image, we calculate and denote the area value of their receptive field as AiLRA_i^{LR} and AjHRA_j^{HR}.

The transformation process in the aforementioned process is implemented by the telescope calibration information carried by the high-resolution (HR) image, which could be interpreted as camera intrinsic and extrinsic parameters of the telescope.

HR Image Subdivision and Data Splitting

The previous flux-consistent downsampling scheme provides flux-consistent HR-LR image pairs. To further optimize model training effectiveness, we divide the HR and LR wide field images into smaller sub-images. Because astronomical images contain some outlier regions (have NaN values) due to geometric calibration in DrizzlePac [51] processing, we retain only patches with >80% valid regions containing stellar features.

To prevent data leakage, we split the dataset by sky region, ensuring that patches from the same sky region are assigned to the same split. We further verify that the training, validation, and test sets have no overlapping sky coverage or shared raw exposures. Through these processing steps, we obtain a large set of high-quality HR-LR image pairs and construct the STAR++ dataset for training and evaluating astronomical super-resolution models.

Flux Error

STAR++ provides a Flux Error (FE) to evaluate flux consistency between a ground truth and its corresponding prediction. The FE measures the flux value gap for each object, so its computation process is based on astrophysical photometry. The basic process of the photometry is deriving flux by detection. We follow this idea. For a given ground-truth and predicted image pair, we compute the FE as following two steps: 1). For the ground-truth image, we use the Starfinder toolkit [52] to detect celestial objects and obtain their parameters. Then, we derive the flux of each object by a widely-used elliptical photometry method [52]. 2). For predicted images, we do not operate object detection but directly use detection results from the ground-truth image because they provide reliable object catalogs covering both strong and weak sources. The following photometry is the same as the ground truth. After these two steps, we have a two set of flux values, denoted as {vpred1,vpred2,…,vpredN}\{v_{\mathrm{pred}}^{1},v_{\mathrm{pred}}^{2},\ldots,v_{\mathrm{pred}}^{N}\} and {vgt1,vgt2,…,vgtN}\{v_{\mathrm{gt}}^{1},v_{\mathrm{gt}}^{2},\ldots,v_{\mathrm{gt}}^{N}\}, where NN is the number of detected objects. The FE is computed by the following:

FE=1N∑i=1N∣vgti−vpredi∣.(3)\mathrm{FE}=\frac{1}{N}\sum_{i=1}^{N}\left|v_{\mathrm{gt}}^{i}-v_{\mathrm{pred}}^{i}\right|. \tag*{(3)}

Lower FE means higher flux consistency. Since the flux value is related to the object shape and the pixel flux in object regions, this metric could reflect the geometric shape consistency for each object in the reconstruction image, physically informed flux consistency and weak source object reconstruction quality simultaneously.

Method

Overview

The overall pipeline of our Flux-Invariant Super-Resolution Plus (FISR++) is shown in Figure 3. The low-resolution input image is first mapped to the asinh intensity space and then processed through two parallel paths, followed by a reconstruction stage. The first one is the Flux Guidance Generation (FGG) module of FISR [19], which extracts flux information as multi-scale flux guidance representations. The second is an encoder block consisting of a convolution operation and multiple transformer blocks, which represent the input image as multi-scale feature maps. Each flux guidance is sent into a scale-wise Flux Guidance Controller (FGC) to enhance the encoded feature maps. This scheme highlights regions with significant flux to guide the network’s focus toward astrophysically relevant structures. These features are then processed by a Filter-Aware Adapter (FAA) to incorporate filter-related physical priors for cross-filter reconstruction. Subsequently, multiple decoder transformer blocks progressively process these features and finally upsample them into the output image. During training, we employ the proposed Spatial Ambiguity Disentangle Loss (SADL) to improve reconstruction robustness.

Figure 3. Overview of the proposed FISR++ framework. The LR input is first transformed into the asinh intensity space and processed by an encoder and a Flux Guidance Generation (FGG) module. The extracted flux guidance is incorporated into multi-scale features through Flux Guidance Controllers (FGCs), while the Filter-Aware Adapter (FAA) integrates physical observation priors to enhance feature representations. The enhanced features are progressively decoded and upsampled to reconstruct the HR image. During training, a pixel-wise uncertainty map is also predicted for the Spatial Ambiguity Disentangle Loss (SADL).

Asinh Intensity Space

We adopt an inverse hyperbolic sine (asinh) transformation to represent astronomical images in a compressed intensity space, inspired by its applications in astronomical photometry and image visualization [53], [54]. Specifically, we transform both the LR inputs into the asinh intensity space for model training and optimization. For an input image IlrI_{\mathrm{lr}}, the transformation is defined as:

Ilr=asinh⁡(Ilr′/β)asinh⁡(1/β),(4)I_{\mathrm{lr}} = \frac{\operatorname{asinh}(I'_{\mathrm{lr}}/\beta)}{\operatorname{asinh}(1/\beta)}, \tag*{(4)}

where asinh⁡(x)=ln⁡(x+x2+1)\operatorname{asinh}(x) = \ln\left(x+\sqrt{x^2+1}\right) is the inverse hyperbolic sine function, and β\beta controls the degree of intensity compression, we set it to 0.1. As a result, the transformed input IlrI_{\mathrm{lr}} enhances the relative intensities of faint stellar sources, reducing the dominance of bright stars in pixel-wise supervision and increasing the contribution of faint sources during training.

Flux Guidance Generation

We propose the (Flux Guidance Generation) FGG module to introduce the flux information. Specifically, given an input image, FGG first represents flux information for every celestial object in the input image into a flux map, then transfers the flux map as multi-scale features as guidance which will then be used in the next FGC module to enhance multi-scale feature maps.

To generate the flux map, FGG also computes flux by detection. It detects celestial objects first and obtains a bounding box for each object. Then the photometry process is operated to derive object fluxes. The entire detection and photometry process is real-time. With these, FGG generates the map following the idea of drawing bounding boxes on a white background and corresponding flux value in each bounding box region.

In practice, because the ’bounding box’ of a celestial object is essentially a rotatable ellipse rather than a rectangle in standard object detection, the drawing scheme FGG has some modifications. Specifically, for each ellipse, FGG puts a rotatable Gaussian Kernel at the center location and adjusts the Gaussian standard deviation based on the ellipse size. Then, FGG multiplies the object flux value directly with the rotated Gaussian Kernel to modulate the kernel value. Finally, FGG draws all these kernels together as the final flux map.

Finally, a multi-scale convolutional block transforms the flux map into a corresponding flux guidance representation. In our architecture, we employ three such blocks at different feature levels to generate a hierarchy of flux guidance, denoted as {G1,G2,G3}\{G_1, G_2, G_3\}. The level 3 is set to align the block number in the decoder, following common setting in a popular SR baseline [55].

Flux Guidance Controller

Based on the flux guidance produced by the FGG module, the (Flux Guidance Controller) FGC interacts such guidance with the encoder features, generating the enhanced features. The FGC is scale-wised and for the ii-th scale FGC, its interaction pipeline is represented as follows:

F^i=PIM⁡i(GCM⁡i(Pi,Fi,Gi),Fi).(5)\hat{F}_i = \operatorname{PIM}_i\left(\operatorname{GCM}_i\left(P_i,F_i,G_i\right),F_i\right). \tag*{(5)}

It shows that the enhanced feature F^i∈RH×W×C\hat{F}_i \in\mathbb{R}^{H\times W\times C} is derived from a guidance-controlled feature and an original feature Fi∈RH×W×CF_i \in\mathbb{R}^{H\times W\times C} by a prompt interaction module function PIM⁡(⋅,⋅)\operatorname{PIM}(\cdot,\cdot), which is set same as combines image features with guidance-controlled feature and dynamically adjusts the input features through a transformer block, following PromptIR [55]. The guidance-controlled features are computed by a guidance controller module function GCM⁡(⋅,⋅,⋅)\operatorname{GCM}(\cdot,\cdot,\cdot) that takes a learnable prompt Pi\mathbf{P}_i, the flux guidance component Gi∈RH×W×C\mathbf{G}_i \in\mathbb{R}^{H \times W \times C} and the encoded feature Fi\mathbf{F}_i as inputs. The learnable prompt Pi∈RH×W×C×K\mathbf{P}_i \in\mathbb{R}^{H \times W \times C \times K} contains KK learnable patterns expected to represent blind property in the image restoration process [55]. The detail of GCM is as follows:

GCM⁡i(Pi,Fi,Gi)=∑k∈KWiF(Fi)⊙WiG(Gi)⊙Pi,(6)\operatorname{GCM}_i(\mathbf{P}_i,\mathbf{F}_i,\mathbf{G}_i)=\sum_{k\in K}W_i^F(\mathbf{F}_i)\odot W_i^G(\mathbf{G}_i)\odot\mathbf{P}_i, \tag*{(6)}

where WiW_i means learnable modules, consisting of a global average pooling and a linear layer, takes input corresponding features and derives a weight with the size of 1×1×1×K1 \times1 \times1 \times K to indicate the importance of the learnable prompt patterns. ⊙\odot means Hadamard product with dimension broadcast while ∑k∈K\sum_{k\in K} means sum the last dimension of the multiplied features with the size of H×W×C×KH \times W \times C \times K to derive the final output features with the size of H×W×CH \times W \times C.

Filter-Aware Adapter

We propose a Filter-Aware Adapter (FAA) to incorporate continuous physical priors into image features for cross-filter reconstruction. Specifically, we obtain filter-related physical parameters mf\mathbf{m}_f from the observation metadata, where meta⁡f\operatorname{meta}_f includes wavelength, bandwidth, and axis ratios. We then normalize these parameters and encode them into filter tokens Pf\mathbf{P}_f through an MLP.

Given the enhanced feature F^i\hat{\mathbf{F}}_i from FGC and the filter tokens Pf\mathbf{P}_f, we employ a multi-head cross-attention mechanism to incorporate the physical priors into image features at different scales. The image features serve as queries, while the filter tokens provide keys and values:

Zi=Softmax⁡(Wq(F^i)Wk(Pf)Td)Wv(Pf),(7)\mathbf{Z}_i=\operatorname{Softmax}\left(\frac{W_q(\hat{\mathbf{F}}_i)W_k(\mathbf{P}_f)^{T}}{\sqrt{d}}\right)W_v(\mathbf{P}_f), \tag*{(7)}

where Wq(⋅)W_q(\cdot), Wk(⋅)W_k(\cdot), and Wv(⋅)W_v(\cdot) denote learnable projection layers for query, key, and value, respectively, and dd is the key dimension. The attended features are projected back to the original channel dimension and fused with F^i\hat{\mathbf{F}}_i through a residual connection:

Fi′=F^i+Wo(Zi),(8)\mathbf{F}'_i=\hat{\mathbf{F}}_i+W_o(\mathbf{Z}_i), \tag*{(8)}

where Wo(⋅)W_o(\cdot) denotes the learnable output projection layer. Since different filters exhibit distinct spectral characteristics and spatial sampling conditions, incorporating these physical priors enables FISR++ to adapt its image representations to different observations, improving generalization to unseen filters.

Spatial Ambiguity Disentangle Loss

To model spatial ambiguity in the reconstruction of high-frequency stellar structures, we propose a Spatial Ambiguity Disentanglement Loss (SADL), as shown in Figure 4. Given the reconstructed image y′∈RH×Wy' \in\mathbb{R}^{H \times W}, the HR ground truth y∈RH×Wy \in\mathbb{R}^{H \times W}, and the predicted pixel-wise uncertainty σ∈RH×W\sigma\in\mathbb{R}^{H \times W}, SADL is formulated as:

Figure 4. Illustration of the Spatial Ambiguity Disentanglement Loss (SADL). During training, SADL learns a global shift δ\delta under a Gaussian prior and applies a first-order Taylor correction to the prediction. The curves illustrate pixel-value profiles, where the vertical distances between the prediction and ground truth represent pixel-wise L1 errors before and after shift correction, weighted by the predicted uncertainty.

LSADL=1HW[∑v=0H−1∑u=0W−1(2σ(u,v)∣y(u,v)−y′(u+Δu,v+Δv)∣)+log⁡σ(u,v)+∥δ∥22σ~2],(9)\mathcal{L}_{\mathrm{SADL}}=\frac{1}{HW}\left[\sum_{v=0}^{H-1}\sum_{u=0}^{W-1}\left(\frac{\sqrt{2}}{\sigma(u,v)}\left|y(u,v)-y'(u+\Delta u,v+\Delta v)\right|\right)+\log\sigma(u,v)+\frac{\lVert\boldsymbol{\delta}\rVert^2}{2\tilde{\sigma}^{2}}\right], \tag*{(9)}

where δ=(Δu,Δv)\boldsymbol{\delta}=(\Delta u,\Delta v). (u,v)(u,v) denotes the pixel coordinates and δ=(Δu,Δv)\boldsymbol{\delta}=(\Delta u,\Delta v) is a learnable global shift pair. The coefficient 2/σ\sqrt{2}/\sigma before the L1 error term serves as an uncertainty-dependent weight. Specifically, when the predicted uncertainty σ(u,v)\sigma(u,v) is large, the coefficient 2/σ(u,v)\sqrt{2}/\sigma(u,v) becomes smaller, reducing the contribution of the corresponding reconstruction error. Meanwhile, the uncertainty regularization term penalizes excessively large uncertainty estimates, preventing the model from trivially reducing the reconstruction error by unboundedly increasing the predicted uncertainty. The spatial shift regularization term plays a similar role to δ\delta, whose regularization strength is controlled by a hyperparameter σ~\tilde{\sigma}. Based on this, to make y′(u+Δu,v+Δv)y'(u+\Delta u,v+\Delta v) differentiable with respect to both y′y' and δ\delta, we approximate it via its first-order Taylor expansion:

y′(u+Δu,v+Δv)≈y′(u,v)+∇y′(u,v)Tδ.(10)y'(u+\Delta u,v+\Delta v)\approx y'(u,v)+\nabla y'(u,v)^{T}\boldsymbol{\delta}. \tag*{(10)}

The first-order term ∇y′(u,v)Tδ\nabla y'(u,v)^{T}\boldsymbol{\delta} approximates the intensity variation induced by the global shift.

The entire SADL is theoretically derived from a maximum a posteriori (MAP) idea to optimize pixel shift, which could be formulated as:

p(δ∣y,y′,σ)=p(y∣y′,δ,σ)p(δ)p(y∣y′,σ)p(\delta\mid y,y',\sigma)=\frac{p(y\mid y',\delta,\sigma)p(\delta)}{p(y\mid y',\sigma)}
∝p(y∣y′,δ,σ)p(δ),(11)\propto p(y\mid y',\delta,\sigma)p(\delta), \tag*{(11)}

The denominator p(y∣y′,σ)p(y\mid y',\sigma) can be omitted during optimization, due to the approximately uniform or slowly varying sky backgrounds commonly observed in local regions of

astronomical images [56, 57]. So, the maximization could be re-written under log space as:

max⁡log⁡p(δ∣y,y′,σ)=max⁡log⁡p(y∣yδ′,σ)+log⁡p(δ),(12)\max\log p(\delta\mid y,y',\sigma)=\max\log p(y\mid y'_{\delta},\sigma)+\log p(\delta), \tag*{(12)}

where p(y∣yδ′,σ)p(y\mid y'_{\delta},\sigma) is modeled by our reconstruction network and p(δ)p(\delta) is assumed as a zero-mean Gaussian prior, δ∼N(0,σ~2I2)\delta\sim\mathcal{N}(0,\tilde{\sigma}^{2}I_{2}). It transfers to the spatial shift regularization term ∥δ∥22/(2σ~2)\lVert\delta\rVert_{2}^{2}/(2\tilde{\sigma}^{2}) constrains the shift magnitude and discourages extreme displacements.

By jointly accounting for pixel-wise uncertainty and latent sub-pixel shifts, SADL reduces sensitivity to positional deviations while preserving the reconstruction constraints on stellar structures. The shift is estimated only during training and is not applied to the reconstructed image at inference. Further implementation details are provided in the supplementary material.

Flux Consistency Loss

The idea of the Flux Consistency Loss is to train the model to generate flux consistent prediction. We propose a simple yet effective scheme to achieve this goal by highlighting flux-related regions. We first operate the flux map generation scheme described in Section IV-C on the ground-truth image, denoted as MM. Additionally, the HR ground-truth image is also transformed into the asinh intensity space to match the reconstructed prediction. They we use this map to weighted the pixel wised supervision as follows:

Lflux(Ipred,Igt)=∑x,yM(x,y)⋅∣Ipred(x,y)−Igt(x,y)∣.(13)\mathcal{L}_{\mathrm{flux}}(I_{\mathrm{pred}},I_{\mathrm{gt}})=\sum_{x,y}M(x,y)\cdot\lvert I_{\mathrm{pred}}(x,y)-I_{\mathrm{gt}}(x,y)\rvert. \tag*{(13)}

Combined with the proposed Spatial Ambiguity Disentangle Loss, the overall training objective is defined as:

Ltotal=LSADL+λfluxLflux,(14)\mathcal{L}_{\mathrm{total}}=\mathcal{L}_{\mathrm{SADL}}+\lambda_{\mathrm{flux}}\mathcal{L}_{\mathrm{flux}}, \tag*{(14)}

where λ\lambda balances terms. This loss takes into account both traditional regression supervision and flux consistency constraints, which brings more physically reliable ASR results.

Experiment

Experimental Setup

Dataset. We consider two training settings, models trained on the original STAR dataset and models trained on the proposed STAR++ dataset. Both are evaluated on the unseen-filter test sets of STAR++ to assess cross-filter generalization and the benefits of multi-filter training.

Implementation Details. All models are trained for 24 epochs with a batch size of 16 on 8 Ascend 910B2 NPUs. The initial learning rate is set to 2×10−42\times10^{-4} and decayed by a factor of 0.01 at the 12th epoch. A linear warm-up strategy is applied during the first epoch to stabilize training. For image degradation, we set the Gaussian PSF parameter σ∈[0.8,1.2]\sigma\in[0.8,1.2] and the Airy PSF radius r∈[1.9,2.2]r\in[1.9,2.2] pixels based on the FWHM [58], which measures the blur width at half its peak intensity, to approximate the actual HST WFC/ACS F814W filter observations where the blur is characterized by its FWHM. The standard deviation of the Gaussian prior on the pixel shift is set to σ~=0.15\tilde{\sigma}=0.15.

Evaluation Protocols. Following prior work, we adopt Peak Signal-to-Noise Ratio (PSNR) [59] and Structural Similarity Index (SSIM) [60] as standard metrics for evaluating reconstruction quality. In addition, we use the proposed FE to evaluate the physical fidelity of the reconstructed images by measuring their consistency with astrophysical properties. We do not adopt commonly used perceptual metrics for image super-resolution, such as LPIPS [61], since these metrics rely on deep models pretrained on natural-image datasets such as ImageNet [62, 63, 64]. Due to the substantial domain gap between natural and astronomical imagery, such pretrained representations may not provide an appropriate measure of perceptual similarity for astronomical images.

Quantitative and Qualitative Results

We compare FISR++ with Bicubic interpolation and representative state-of-the-art SR methods, including CNN-based EDSR [8] and RCAN, Transformer-based SwinIR [28] and HAT [29], and the astronomical SR method FISR [19] are also popular in the ASR topic [11, 18, 22], which is also an important factor that we choose them to compare. As shown in Table III and Table IV, we compare models trained on STAR and STAR++ using the same unseen-filter test sets of STAR++, and the results show that our approach outperforms existing methods across different evaluation criteria.

Training setMethodPSNR (dB) ↑ Interp.PSNR (dB) ↑ Extrap.PSNR (dB) ↑ OverallSSIM ↑FE ↓
–Bicubic29.7223.5327.060.75852.8908
STAREDSR31.9424.1628.610.79341.2957
STARRCAN33.8626.7930.830.82010.7404
STARSwinIR34.7926.9531.430.82640.7141
STARHAT35.1426.7631.550.82530.6205
STARFISR35.5727.3432.040.82160.6027
STARFISR++36.8928.6333.350.82800.5888
STAR++EDSR36.0829.3933.210.83910.7846
STAR++RCAN36.7430.7334.160.84140.7247
STAR++SwinIR37.1230.3534.220.85220.6607
STAR++HAT37.9331.2835.080.85850.5617
STAR++FISR37.2230.4634.320.85420.5304
STAR++FISR++38.0531.9135.420.86230.4341

Table III. ASR PERFORMANCE ON SEVEN UNSEEN STAR++ FILTERS AT ×2. THE INTERPOLATION SUBSET COMPRISES F435W, F555W, F606W, AND F625W, WHILE THE EXTRAPOLATION SUBSET COMPRISES F218W, F225W, AND F850LP. STAR AND STAR++ INDICATE THE TRAINING DATASETS.

Training setMethodPSNR (dB) ↑ Interp.PSNR (dB) ↑ Extrap.PSNR (dB) ↑ OverallSSIM ↑FE ↓
–Bicubic26.2320.4123.740.65234.4267
STAREDSR30.8423.4427.670.71651.0674
STARRCAN31.5323.6728.160.72340.8701
STARSwinIR31.6824.0328.400.72830.8273
STARHAT31.9924.0328.580.73280.7659
STARFISR32.2424.1128.760.73450.7670
STARFISR++32.5325.7529.620.76530.7540
STAR++EDSR33.0826.9230.440.75730.8683
STAR++RCAN33.6926.8330.750.76340.7639
STAR++SwinIR34.0327.3131.150.76810.7559
STAR++HAT34.4427.7631.580.77730.7494
STAR++FISR34.0928.1031.520.77360.7257
STAR++FISR++35.3529.0132.630.78050.6803

Table IV. ASR PERFORMANCE ON UNSEEN STAR++ FILTERS AT ×4. STAR AND STAR++ INDICATE THE TRAINING DATASETS.

Compared with their STAR-trained counterparts, models trained on STAR++ generally achieve higher overall PSNR on unseen filters. Specifically, EDSR improves from 28.61 to 33.21 dB at ×2\times2 and from 27.67 to 30.44 dB at ×4\times4. HAT achieves gains of 3.53 dB and 3.00 dB at the two scales, respectively, while FISR improves by 2.28 dB and 2.76 dB.

We further analyze generalization to unseen interpolation and extrapolation filters. At ×2\times2, FISR++ achieves PSNR values of 38.05 dB and 31.91 dB on the interpolation and extrapolation subsets, respectively, outperforming FISR by 0.83 dB and 1.45 dB. At ×4\times4, the corresponding values reach 35.35 dB and 29.01 dB, with improvements of 1.26 dB and 0.91 dB over FISR. Overall, these results indicate that training on STAR++ generally improves model generalization to unseen filters, while FISR++ consistently outperforms FISR across both interpolation and extrapolation settings, demonstrating its superior cross-filter generalization capability.

Among models trained on STAR++, FISR++ achieves the highest overall PSNR of 35.42 dB and 32.63 dB at ×2\times2 and ×4\times4, respectively. Compared with FISR under the same training setting, FISR++ improves PSNR by 1.10 dB and 1.11 dB, while reducing FE from 0.5304 to 0.4341 and from 0.7257 to 0.6803 at the two scales, respectively. These results demonstrate the effectiveness of FISR++ in improving reconstruction accuracy and flux consistency across unseen filters.

Ablation Study

To evaluate the contribution of each proposed component in our FISR++ framework, we conduct comprehensive ablation studies on the STAR++ dataset under the ×2\times2 super-resolution settings. Tab. V reports the ablation results of the three conference-version components (FGG, FGC, and FCL) on STAR, followed by the three newly introduced components (FAA, SADL, and Asinh) on STAR++. Specifically, SADL is further decomposed into uncertainty loss (UL) and spatial shift regularization (SSR) to evaluate their individual contributions. We also investigate the sensitivity of SADL to the Gaussian prior standard deviation σe\sigma_e, as reported in Tab. VI. We further evaluate the effectiveness of FCL across seven model configurations at ×2 and ×4 scales. As shown in Tab. VII, incorporating FCL consistently reduces FE across all seven configurations at both scales. At ×2, FE decreases from 0.5775 to 0.5585 for FISR trained on STAR and from 0.4410 to 0.4341 for FISR++ trained on STAR++. At ×4, the corresponding reductions are from 1.0861 to 0.9625 for FISR and from 0.6811 to 0.6803 for FISR++. These results demonstrate that FCL consistently improves photometric fidelity in terms of FE, although the magnitude of improvement varies across models and upsampling scales.

Effect of FGG, FGC, and FCL: We briefly revisit the three components introduced in the conference version. As shown in configurations (a)–(f), although individual combinations do not consistently improve all metrics, integrating all three components achieves the best overall performance on STAR, improving PSNR from 37.77 to 37.88 dB and reducing FE from 0.7022 to 0.5739. These results confirm the combined effectiveness of the original flux-guidance design.

Effect of FAA: Comparing configurations (g) and (h), introducing FAA improves PSNR by 0.32 dB and SSIM by 0.0029, while reducing FE from 0.5304 to 0.5207. These results demonstrate the effectiveness of incorporating filter-specific physical priors into the reconstruction framework.

Effect of SADL: Compared with configuration (g), introducing UL in (i) improves PSNR by 0.27 dB and SSIM by 0.0032, while reducing FE from 0.5304 to 0.4714. Further incorporating spatial shift modeling with regularization in (j) improves PSNR by an additional 0.28 dB and SSIM by 0.0029, while reducing FE to 0.4549. Moreover, jointly incorporating FAA and the complete SADL in (k) improves PSNR by 0.95 dB and SSIM by 0.0072, while reducing FE to 0.4536 compared with (g). These results demonstrate the complementary benefits of pixel-wise uncertainty modeling and spatial shift regularization, with further improvements from filter-aware adaptation.

DatasetIDFGGFGCFCLFAAULSSRAsinhPSNR↑SSIM↑FE↓
STAR(a)-------37.770.82880.7022
STAR(b)--✓----37.860.82830.6389
STAR(c)✓------37.710.82860.7572
STAR(d)✓-✓----37.870.83010.7467
STAR(e)✓✓-----37.850.83020.6527
STAR(f)✓✓✓----37.880.83110.5739
STAR++(g)✓✓✓----34.320.85420.5304
STAR++(h)✓✓✓✓---34.640.85710.5207
STAR++(i)✓✓✓-✓--34.590.85740.4714
STAR++(j)✓✓✓-✓✓-34.870.86030.4549
STAR++(k)✓✓✓✓✓✓-35.270.86140.4536
STAR++(l)✓✓✓✓✓✓✓35.420.86230.4341

Table 5. Ablation studies of different components under the ×2 super-resolution setting. UL and SSR denote Uncertainty Loss and Spatial Shift Regularization, respectively.

σ~\tilde{\sigma}PSNR (dB)↑SSIM↑FE↓
0.0535.370.86180.4378
0.1035.390.86190.4347
0.1535.420.86230.4341
0.2035.410.86220.4343
0.5034.940.86070.4401

Table 6. Sensitivity analysis of the Gaussian prior standard deviation σ~\tilde{\sigma} in SADL. The best results are highlighted in bold.

Effect of Asinh: Comparing configurations (k) and (l), introducing the asinh intensity representation improves PSNR by 0.15 dB and SSIM by 0.0009, while reducing FE from 0.4536 to 0.4341. The complete FISR++ achieves the best overall performance, outperforming the FISR baseline (g) by 1.10 dB in PSNR. These results suggest that compressing the intensity dynamic range helps balance the contributions of bright and faint sources during reconstruction, improving both reconstruction quality and photometric fidelity.

Sensitivity Analysis of the Shift Prior: We investigate the sensitivity of SADL to the prior standard deviation σ~\tilde{\sigma} by varying it from 0.05 to 0.50. As shown in Table VI, the reconstruction performance remains relatively stable when σ~\tilde{\sigma} ranges from 0.10 to 0.20. The best performance is achieved at σ~=0.15\tilde{\sigma}=0.15, with an PSNR of 35.42 dB, an SSIM of 0.8623, and an FE of 0.4341. When σ~\tilde{\sigma} increases to 0.50, PSNR decreases by 0.48 dB, indicating that excessively weak shift regularization may degrade reconstruction performance. These results support the choice of σ~=0.15\tilde{\sigma}=0.15 as the default setting.

ScaleFlux LossEDSR [8]RCAN [24]SwinIR [28]HAT [29]FISRFISR++FISR++
×2PSNR↑ (w/o)35.7437.4637.6037.9138.0138.0935.28
SSIM↑ (w/o)0.80860.82770.82790.82960.83130.83460.8605
FE↓ (w/o)1.32490.74510.75930.69180.57750.54690.4410
PSNR↑ (w/)35.8937.8937.1337.9438.2738.8335.42
SSIM↑ (w/)0.80920.82860.82810.83030.83340.83840.8623
FE↓ (w/)1.24200.59140.67670.57380.55850.50120.4341
×4PSNR↑ (w/o)33.2135.7035.1135.8736.2336.9132.47
SSIM↑ (w/o)0.72010.72630.72580.72760.72330.76280.7765
FE↓ (w/o)1.6811.23551.37051.11631.08610.99480.6811
PSNR↑ (w/)33.4535.4335.3236.1636.4637.1732.63
SSIM↑ (w/)0.72950.73100.73380.73640.73520.76620.7805
FE↓ (w/)1.24591.00550.92340.92990.96250.68030.6803

Table VII. Comparison of different methods with and without FCL under ×2 and ×4 ASR settings on STAR and STAR++. Bold entries indicate that the setting with FCL (w/) improves over the setting without FCL (w/o) for the same metric and method.

Quality Analysis

Figure 5 and Figure 6 present qualitative comparisons of different SR methods on unseen interpolation and extrapolation filters, respectively. Although existing methods can recover the overall structures of star fields, they often produce noticeable reconstruction residuals, particularly around bright stars and densely populated regions. In contrast, FISR++ produces cleaner residual maps across both settings while preserving stellar structures, demonstrating its effectiveness in reconstructing unseen-filter observations.

Figure 5. Visualization of SR reconstruction results on unseen interpolation filters. For each filter, the upper row shows the reconstructions from different methods, while the lower row presents the corresponding residual maps.

Figure 6. Visualization of SR reconstruction results on unseen extrapolation filters. For each filter, the upper row shows the reconstructions from different methods, while the lower row presents the corresponding residual maps.

To further evaluate pixel shift in reconstructed star fields, Figure 7 presents the stellar centroid offsets obtained by different methods. The cyan crosses indicate the ground-truth centroids, while the red lines illustrate the positional deviations. FISR++ achieves smaller centroid offsets than the compared methods for all six selected sources. In particular, its offsets are consistently smaller than those of FISR, suggesting that the proposed SADL helps reduce spatial misalignment during reconstruction. These results provide further evidence that FISR++ not only improves pixel-level reconstruction quality but also better preserves the positional accuracy of individual stellar sources.

Visualization of stellar centroid shifts under different SR methods

Figure 7. Visualization of stellar centroid shifts under different SR methods. Left: six selected sources marked by boxes. Right: comparison of their centroid offsets across different SR methods.

Conclusion

In this work, we present STAR++, a large-scale, multi-filter benchmark for astronomical super-resolution (ASR). To address the limited filter diversity and lack of systematic unseen-filter evaluation in existing benchmarks, STAR++ provides approximately 1.7 million flux-consistent LR–HR image pairs across 36 HST filters, with sky-region-disjoint splits supporting the evaluation of cross-filter generalization under both interpolation and extrapolation settings. Building on STAR++, we further propose FISR++, which extends the original flux-guided reconstruction framework with a Filter-Aware Adapter (FAA), an asinh intensity representation, and a Spatial Ambiguity Disentangle Loss (SADL). Specifically, these components incorporate filter-specific physical information into feature representations, balance the contributions of bright and faint sources during training, and explicitly account for pixel-wise reconstruction uncertainty and sub-pixel spatial misalignment, respectively. Extensive experiments demonstrate that training on STAR++ consistently improves unseen-filter reconstruction across multiple SR architectures. Furthermore, FISR++ achieves the best overall performance among the compared methods, attaining the highest PSNR and SSIM and the lowest Flux Error (FE) at both ×2\times2 and ×4\times4 super-resolution scales.

References

  1. [1]A Labeyrie. Hi high-resolution techniques in optical astronomy. Progress in optics, 14:47–87, 1977.
  2. [2]Aaron Bryant and Alfred Krabbe. The episodic and multiscale galactic centre. New Astronomy Reviews, 93:101630, 2021.
  3. [3]Muhammad Faique. Overview of big data analytics in modern astronomy. International Journal of Mathematics, Statistics, and Computer Science, 2:96–113, 2024.DOI
  4. [4]W Patrick McCray. Giant telescopes: Astronomical ambition and the promise of technology. Harvard University Press, 2006.
  5. [5]W Patrick McCray. Large telescopes and the moral economy of recent astronomy. Social studies of science, 30(5):685–711, 2000.DOI
  6. [6]Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295–307, 2015.
  7. [7]Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197–5206, 2015.
  8. [8]Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136–144, 2017.
  9. [9]R Marsh, TR Young, T Johnson, and D Smith. Enhancement of small telescope images using super-resolution techniques. Publications of the Astronomical Society of the Pacific, 116(819):477, 2004.
  10. [10]Klaus G Puschmann and Franz Kneer. On super-resolution in astronomical imaging. Astronomy & Astrophysics, 436(1):373–378, 2005.
  11. [11]Jiawei Miao, Liangping Tu, Bin Jiang, Xiangru Li, and Bo Qiu. Astrosr: A data set of galaxy images for astronomical superresolution research. The Astrophysical Journal Supplement Series, 274(1):7, 2024.DOI
  12. [12]Qian-Qian Shan, Cui-Xiang Liu, Bo Qiu, A-Li Luo, Fu-ji Ren, Zhi-Ren Pan, and Yi-Cong Chen. Galaxy image super-resolution reconstruction using diffusion network. Engineering Applications of Artificial Intelligence, 142:109836, 2025.
  13. [13]Zhijian Luo, Shaohua Zhang, Jianzhen Chen, Zhu Chen, Liping Fu, Hubing Xiao, Wei Du, and Chenggang Shu. Cross-survey image transformation: Enhancing sdss and decals images to near-hsc quality for advanced astronomical analysis. The Astrophysical Journal Supplement Series, 277(1):22, 2025.
  14. [14]Karl D Gordon, Ralph Bohlin, GC Sloan, George Rieke, Kevin Volk, Martha Boyer, James Muzerolle, Everett Schlawin, Susana E Deustua, Dean C Hines, et al. The james webb space telescope absolute flux calibration. i. program design and calibrator stars. The Astronomical Journal, 163(6):267, 2022.DOI
  15. [15]Allan W Smith, John T Woodward, Colleen A Jenkins, Steven W Brown, and Keith R Lykke. Absolute flux calibration of stars: calibration of the reference telescope. Metrologia, 46(4):S219, 2009.
  16. [16]Ralph C Bohlin, Susana E Deustua, and Gisella De Rosa. Hubble space telescope flux calibration. i. stis and calspec. The Astronomical Journal, 158(5):211, 2019.DOI
  17. [17]Wei Song, Ying Ma, Haoying Sun, Xiaobing Zhao, and Ganghua Lin. Improving the spatial resolution of solar images using super-resolution diffusion generative adversarial networks. Astronomy & Astrophysics, 686:A272, 2024.
  18. [18]Pranath Reddy, Michael W Toomey, Hanna Parul, and Sergei Gleyzer. Defense: a conditional diffusion model for super-resolution of gravitational lensing data. Machine Learning: Science and Technology, 5(3):035076, 2024.arxiv.org/abs/2406.08442
  19. [19]WU KUO-CHENG, Guohang Zhuang, Jinyang Huang, Xiang Zhang, Wanli Ouyang, and Yan Lu. Star: A benchmark for astronomical star fields super-resolution. Advances in Neural Information Processing Systems, 38, 2026.
  20. [20]Wanjun Li, Zhe Liu, and Hongtao Deng. A self-attention based srgan for super-resolution of astronomical image. In 2022 IEEE 8th International Conference on Computer and Communications (ICCC), pages 1977–1981. IEEE, 2022.
  21. [21]Wanli Ouyang, Yan Lu, Hao Du, Jiaze Li, Kuo-cheng Wu, Zhen Wan, Guohang Zhuang, Yating Liu, Shixiang Tang, Lei Bai, et al. Imaging physics-driven artificial intelligence makes ground-based telescope resolve deep field universe. 2025.DOI
  22. [22]Shuhong Liu, Xining Ge, Zitong Cui, Liuzhouzheng Li, Gengjia Chang, Jun Liu, Ziying Gu, Dong Li, Xuangen Chu, Lin Gu, et al. Fluxflow: Conservative flow-matching for astronomical image super-resolution. arXiv preprint arXiv:2605.03749, 2026.arxiv.org/abs/2605.03749
  23. [23]Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711–730. Springer, 2010.
  24. [24]Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286–301, 2018.
  25. [25]Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro A. A. Este, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681–4690, 2017.
  26. [26]Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905–1914, 2021.arxiv.org/abs/2107.10833
  27. [27]Kelvin CK Chan, Xiangyu Xu, Xintao Wang, Jinwei Gu, and Chen Change Loy. Generalize latent bank for image super-resolution and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3154–3168, 2022.
  28. [28]Jingyun Liang, Jingyun Gao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833–1844, 2021.
  29. [29]Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super-resolution transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2267–2277, 2023.
  30. [30]Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image super-resolution. In European conference on computer vision, pages 649–667. Springer, 2022.arxiv.org/abs/2203.06697
  31. [31]Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457–466, 2022.
  32. [32]Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713–4726, 2022.
  33. [33]Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 132(12):5929–5949, 2024.
  34. [34]Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022.
  35. [35]Jean-Luc Starck, Eric Pantin, and Fionn Murtagh. Deconvolution in astronomy: A review. Publications of the Astronomical Society of the Pacific, 114(800):1051, 2002.DOI
  36. [36]James B Breckinridge, Wai Sze T Lam, and Russell A Chipman. Polarization aberrations in astronomical telescopes: the point spread function. Publications of the Astronomical Society of the Pacific, 127(951):445, 2015.DOI
  37. [37]Michael Hirsch, S Harmeling, S Sra, and B Schölkopf. Online multi-frame blind deconvolution with super-resolution and saturation correction. Astronomy & Astrophysics, 531:A9, 2011.DOI
  38. [38]Stefano Zibetti, Brice Ménard, Daniel B Nestor, Anna M Quider, Sandhya M Rao, and David A Turnshek. Optical properties and spatial distribution of mg ii absorbers from sdss image stacking. The Astrophysical Journal, 658(1):161, 2007.DOI
  39. [39]Kevin Schawinski, Ce Zhang, Hantian Zhang, Lucas Fowler, and Gokula Krishnan Santhanam. Generative adversarial networks recover features in astrophysical images of galaxies beyond the deconvolution limit. Monthly Notices of the Royal Astronomical Society: Letters, 467(1):L110–L114, 2017.DOI
  40. [40]CJ Díaz Baso and A Asensio Ramos. Enhancing sdo/hmi images using deep learning. Astronomy & Astrophysics, 614:A5, 2018.
  41. [41]Sam F Sweere, Ivan Valtchanov, Maggie Lieu, Antonia Vojtě́kova, Eva Verdugo, Maria Santos-Lleo, Florian Pacaud, Alexia Briassouli, and Daniel Cámpora Pérez. Deep learning-based super-resolution and de-noising for xmm-newton images. Monthly Notices of the Royal Astronomical Society, 517(3):4054–4069, 2022.DOI
  42. [42]Sidney Van den Bergh. Ubv photometry of globular clusters. Astronomical Journal, Vol. 72, p. 70-81 (1967), 72:70–81, 1967.
  43. [43]Michael S Bessell. Standard photometric systems. Annu. Rev. Astron. Astrophys., 43(1):293–336, 2005.
  44. [44]Nikhil Padmanabhan, David J Schlegel, Douglas P Finkbeiner, JC Barentine, Michael R Blanton, Howard J Brewington, James E Gunn, Michael Harnavak, David W Hogg, Željko Ivezić, et al. An improved photometric calibration of the sloan digital sky survey imaging data. The Astrophysical Journal, 674(2):1217, 2008.DOI
  45. [45]Holland C Ford, Mark Clampin, George F Hartig, Garth D Illingworth, Marco Sirianni, Andre R Martel, Gerhardt R Meurer, William Jon McCann, Pamela C Sullivan, Frank Bartko, et al. Overview of the advanced camera for surveys on-orbit performance. In Future EUV/UV and Visible Space Astrophysics Missions and Instrumentation, volume 4854, pages 81–94. SPIE, 2003.
  46. [46]L Dressel, M Wong, C Pavlovsky, K Long, et al. Wide field camera 3 instrument handbook, 2010.
  47. [47]Jan J Koenderink and Andrea J van Doorn. Representation of local geometry in the visual system. Biological cybernetics, 55(6):367–375, 1987.
  48. [48]John W Hardy. Adaptive optics for astronomical telescopes, volume 16. Oxford university press, 1998.
  49. [49]François Roddier. V the effects of atmospheric turbulence in optical astronomy. In Progress in optics, volume 19, pages 281–376. Elsevier, 1981.
  50. [50]Nick Scoville, RG Abraham, H Aussel, JE Barnes, A Benson, AW Blain, D Calzetti, A Comastri, C Papak, C Carilli, et al. Cosmos: Hubble space telescope observations. The Astrophysical Journal Supplement Series, 172(1):38, 2007.
  51. [51]S Gonzaga, W Hack, A Fruchter, and J e Mack. The DrizzlePac Handbook, 2012.
  52. [52]Kyle Barbary. Sep: Source extraction and photometry. Astrophysics Source Code Library, pages ascl–1811, 2018.
  53. [53]Robert Lupton, Michael R Blanton, George Fekete, David W Hogg, Wil O’Mullane, Alex Szalay, and Nicholas Wherry. Preparing red-green-blue images from ccd data. Publications of the Astronomical Society of the Pacific, 116(816):133–137, 2004.DOI
  54. [54]Dominic Stark, Barthelemy Launet, Kevin Schawinski, Ce Zhang, Michael Koss, M Dennis Turp, Lia F Sartori, Hantian Zhang, Yiru Chen, and Anna K Weigel. Psfgan: a generative adversarial network system for separating quasar point sources and host galaxy light. Monthly Notices of the Royal Astronomical Society, 477(2):2513–2527, 2018.DOI
  55. [55]Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Promptr: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36:71275–71293, 2023.DOI
  56. [56]Emmanuel Bertin and Stephane Arnouts. Sextractor: Software for source extraction. Astronomy and astrophysics supplement series, 117(2):393–404, 1996.
  57. [57]ASG Robotham, LJM Davies, SP Driver, S Koushan, DS Taranu, S Casura, and J Liske. Profound: source extraction and application to modern survey data. Monthly Notices of the Royal Astronomical Society, 476(3):3137–3159, 2018.DOI
  58. [58]M Sirianni, MJ Jee, N Benítez, JP Blakeslee, AR Martel, Gerhardt Meurer, M Clampin, G De Marchi, HC Ford, R Gilliland, et al. The photometric performance and calibration of the hubble space telescope advanced camera for surveys. Publications of the Astronomical Society of the Pacific, 117(836):1049, 2005.
  59. [59]Jochen Antkowiak, T Jamal Baina, France Vittorio Baroncini, Noel Chateau, France FranceTelecom, Antonio Claudio França Pessoa, F Stephanie Colonnese, Italy Laura Contin, Jorge Caviedes, and France Philips. Final report from the video quality experts group on the validation of objective models of video quality assessment march 2000. Final report from the video quality assessment group on the validation of objective models of video quality assessment march, 2000, 2000.
  60. [60]Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004.
  61. [61]Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018.
  62. [62]Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009.
  63. [63]Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.DOI
  64. [64]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.

Paper details

vlnc:W_qc7bqnyj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents