Research map / Image generation / Super-resolution
STAR: A Benchmark for Astronomical Star Fields Super-Resolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces STAR, a 54,738-pair field-level astronomical super-resolution benchmark from HST F814W star fields with flux-consistent downsampling, a Flux Error metric, and the FISR model with a flux consistency loss.
Problem
Existing astronomical super-resolution datasets are physically inconsistent in flux, crop single objects, and are small in scale, which limits model development and fair evaluation.
Approach
Builds cross-resolution pairs with a flux-conserving (WCS-based, area-weighted) downsampling pipeline on field-level HST images, adds a photometry-based Flux Error metric, and proposes FISR which injects detected-source flux maps as guidance and uses a flux-weighted loss.
Key results
- FISR x2 on STAR reaches PSNR 37.8779, SSIM 0.8311, FE 0.5739, versus HAT at 37.2501 / 0.8295 / 0.7636.
- Adding FCL lowers FE for most baselines in the x2 setting, e.g., SwinIR FE 0.8130 to 0.7809 and RCAN FE 0.9237 to 0.6631.
- Reported FISR is a 24.84% improvement on the flux consistency metric over SOTA methods; under Gaussian+Airy PSF it gives 3.05% higher PSNR and 26.45% lower FE than SwinIR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FISR instead of HAT | STAR ×2 · STAR ×2 | PSNR: 37.2501 -> 37.8779 | Helped |
| Use FISR instead of HAT | STAR ×2 · STAR ×2 | Flux Error: 0.7636 -> 0.5739 | Helped |
| Use FISR instead of HAT | STAR ×4 · STAR ×4 | SSIM: 0.7276 -> 0.7266 | Hurt |
| Add FCL to the PromptIR baseline (ablation) | PromptIR · STAR ×2 | Flux Error: 0.7022 -> 0.6389 | Helped |
| Add FG-Modules without input flux cues (ablation) | PromptIR · STAR ×2 | Flux Error: 0.7022 -> 0.7572 | Hurt |
| Add FG-Modules and FCL, without input flux cues (ablation) | PromptIR · STAR ×2 | Flux Error: 0.7022 -> 0.7467 | Hurt |
| Add input flux cues to FG-Modules (ablation) | PromptIR · STAR ×2 | Flux Error: 0.7572 -> 0.6527 | Helped |
| Add FCL to FG-Modules with input flux cues (ablation) | PromptIR · STAR ×2 | Flux Error: 0.6527 -> 0.5739 | Helped |
| Use Airy instead of Gaussian kernel in FGG (ablation) | FISR · STAR ×2, Gaussian + Airy PSF | PSNR: 37.8779 -> 37.6988 | Hurt |
| Use a random Gaussian/Airy kernel mix in FGG (ablation) | FISR · STAR ×2, Gaussian + Airy PSF | PSNR: 37.8779 -> 37.8186 | No clear effect |
| Use noisy rather than clean source detections in FGG (ablation) | FISR · STAR ×2, Gaussian + Airy PSF | PSNR: 37.8779 -> 37.3176 | Hurt |
| Set FCL weight λ to 0.01 rather than 0.1 (ablation) | FISR · STAR ×2, Gaussian + Airy PSF | PSNR: 37.0843 -> 37.6049 | Helped |
Surprising findings
- Adding the flux-guidance modules without flux cues (only extra parameters) does not improve or even worsens Flux Error, indicating that gains come from flux information rather than capacity.
- Standard perceptual metric LPIPS was deliberately excluded because pretrained ImageNet features have a large domain gap with astronomical images.
Limitations
- Experiments use only a single telescope (HST WFC/ACS) and the F814W filter, limiting generalization.
- Network design lacks explicit astronomical or physical priors beyond the flux guidance.
- Flux Error depends on the detection and photometry pipeline applied to ground truth, so detection errors propagate into the metric.
Experimental setup
Datasets: STAR (HST ACS/WFC F814W) AstroSR
Benchmarks: STAR AstroSR
Metrics: PSNR SSIM Flux Error KL divergence JS divergence recall distance MAE stellar mass MAE mean shear error
Baselines: Bicubic EDSR RCAN SwinIR HAT RealESRGAN PromptIR
Builds on: PromptIR SwinIR DrizzlePac Starfinder
Closest papers by meaning
- FluxFlow: Conservative Flow-Matching for Astronomical Image Super-Resolution (NeurIPS 2026 · similarity 0.88)
- A New Statistical Model of Star Speckles for Learning to Detect and Characterize Exoplanets in Direct Imaging Observations (CVPR 2025 · similarity 0.81)
- AstroCompress: A benchmark dataset for multi-purpose compression of astronomical data (ICLR 2025 · similarity 0.81)
- A systematic evaluation of vision-language models for observational astronomical reasoning tasks (NeurIPS 2026 · similarity 0.80)
- AstroVisBench: A Code Benchmark for Scientific Computing and Visualization in Astronomy (NeurIPS 2025 · similarity 0.79)
- BERTground: A Transformer-Based Model of Background Spectra on the ISS-Based NICER Space Telescope (AAAI 2024 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.