STAR: A Benchmark for Astronomical Star Fields Super-Resolution

TL;DR

STAR addresses flux inconsistency, limited diversity, and object-centric bias in astronomical super-resolution by providing a large-scale, flux-consistent star-field dataset derived from HST, complemented by a Flux Error metric for physical fidelity. It introduces Flux-Invariant Super Resolution (FISR), which integrates flux cues via Flux Guidance Generation and Flux Consistency Loss to preserve photometry while improving resolution. Across extensive experiments, FISR achieves state-of-the-art reconstruction quality (PSNR/SSIM) and superior flux preservation (FE), and yields tangible gains on downstream scientific tasks like stellar mass estimation and weak-lensing measurements. The combination of STAR and FISR enables more reliable scientific analyses in astronomy and lays groundwork for applying SR to wide-field, flux-constrained celestial imaging.

Abstract

Super-resolution (SR) advances astronomical imaging by enabling cost-effective high-resolution capture, crucial for detecting faraway celestial objects and precise structural analysis. However, existing datasets for astronomical SR (ASR) exhibit three critical limitations: flux inconsistency, object-crop setting, and insufficient data diversity, significantly impeding ASR development. We propose STAR, a large-scale astronomical SR dataset containing 54,738 flux-consistent star field image pairs covering wide celestial regions. These pairs combine Hubble Space Telescope high-resolution observations with physically faithful low-resolution counterparts generated through a flux-preserving data generation pipeline, enabling systematic development of field-level ASR models. To further empower the ASR community, STAR provides a novel Flux Error (FE) to evaluate SR models in physical view. Leveraging this benchmark, we propose a Flux-Invariant Super Resolution (FISR) model that could accurately infer the flux-consistent high-resolution images from input photometry, suppressing several SR state-of-the-art methods by 24.84% on a novel designed flux consistency metric, showing the priority of our method for astrophysics. Extensive experiments demonstrate the effectiveness of our proposed method and the value of our dataset. Code and models are available at https://github.com/GuoCheng12/STAR.

Figures (8)

• Figure 1: Comparison of previous datasets and ours, highlighting richer structures such as cross-object interaction, weak lensing, and dark matter halos.
• Figure 2: Overview of the FISR approach. The input image is processed through an encoder branch and a Flux Guidance Generation (FGG). Flux information is extracted via FGG and injected into the Flux Guidance Controllers (FGC) to enhance the encoded feature map. The enhanced features are then decoded and upsampled to produce the final high-resolution output..
• Figure 3: Visual comparison on two star field regions. Red boxes mark areas for computing KL and JS divergence between predictions and ground truth. Models with ($*$) are trained using FCL.
• Figure 4: 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.
• Figure 5: Comparison between downsampling methods. Each HR patch is shown alongside three columns: (a) flux-consistent downsampling, (b) bilinear interpolation, and (c) their pixel-wise difference. FM means flux mean.