Research map / Image generation / Super-resolution
FluxFlow: Conservative Flow-Matching for Astronomical Image 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.
Proposes FluxFlow, a pixel-space optimal-transport flow-matching model for ground-to-space astronomical super-resolution with inverse-variance and source-region weighting plus a training-free Wiener-regularized measurement-consistency correction, and releases the DESI-HST dataset of 19,500 real image pairs.
Problem
Ground-based survey images (DESI, seeing-limited) lack the resolution of HST; existing astronomical SR methods are trained on synthetic degradations and generative models hallucinate compact sources that bias scientific measurements.
Approach
Pixel-space OT-CFM trained with per-pixel inverse-variance weights and source-region weights, plus an inference-time correction that back-projects the LR residual through a Wiener-regularized inverse PSF operator instead of the plain adjoint, avoiding cumulative PSF widening while suppressing hallucinated sources.
Key results
- At x2, FluxFlow achieves SSIM 0.714 and Flux-L1 2.959, best among compared methods, with PSNR 31.23 vs. 31.31 for HAT.
- At x4, FluxFlow achieves SSIM 0.570 (tied best with HAT) and Flux-L1 3.755, best among methods, with PSNR 29.14 vs. 29.21 for HAT.
- Ablation: removing MC-FS drops precision to 0.602 from 0.738 full-model, showing hallucination suppression; removing Wiener correction lowers PSNR to 30.79 from 31.23.
Surprising findings
- Naively applying the physical adjoint back-projection during ODE sampling accumulates PSF widening across steps, motivating the Wiener-regularized correction.
- Removing the WHT weighting yields a marginally lower Flux-L1 (2.814 vs 2.959), which the authors attribute to fitting drizzle artifacts overlapping source apertures.
Limitations
- Hallucinations are suppressed rather than eliminated.
- Scientific accuracy still falls short of what precision astronomical measurements require.
- Single-band (i/F814W) only; no multi-band joint reconstruction.
Experimental setup
Datasets: DESI-HST Dataset
Benchmarks: DESI-HST Dataset AstroSR STAR
Metrics: PSNR SSIM Flux-L1 precision recall F1
Baselines: Bicubic SwinIR HAT FISR cGAN GD-Net AS-Bridge
Builds on: flow matching optimal-transport conditional flow matching diffusion posterior sampling pseudoinverse-guided diffusion models WienerFlow
Closest papers by meaning
- STAR: A Benchmark for Astronomical Star Fields Super-Resolution (NeurIPS 2025 · similarity 0.88)
- Adaptive Flow Matching for Resolving Small-Scale Physics (ICML 2025 · similarity 0.86)
- A New Statistical Model of Star Speckles for Learning to Detect and Characterize Exoplanets in Direct Imaging Observations (CVPR 2025 · similarity 0.84)
- Boosting Flow-based Generative Super-Resolution Models via Learned Prior (CVPR 2024 · similarity 0.82)
- One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation (ICML 2025 · similarity 0.82)
- Neural Eulerian Scene Flow Fields (ICLR 2025 · similarity 0.81)
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.