pith. machine review for the scientific record. sign in

arxiv: 2605.03749 · v2 · submitted 2026-05-05 · 💻 cs.CV

Recognition: 3 theorem links

· Lean Theorem

FluxFlow: Conservative Flow-Matching for Astronomical Image Super-Resolution

Shuhong Liu , Xining Ge , Ziteng Cui , Liuzhuozheng Li , Gengjia Chang , Jun Liu , Ziying Gu , Dong Li
show 3 more authors
Xuangeng Chu Lin Gu Tatsuya Harada
Authors on Pith no claims yet

Pith reviewed 2026-05-08 18:51 UTC · model grok-4.3

classification 💻 cs.CV
keywords astronomical image super-resolutionflow matchingconservative generationatmospheric PSFhallucination suppressionWiener regularizationreal-world benchmarkground-to-space imaging
0
0 comments X

The pith

FluxFlow recovers space-quality astronomical images from ground observations using conservative flow-matching that accounts for uncertainties and suppresses hallucinations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Recovering sharp space-telescope images from ground-based observations is difficult because atmospheric turbulence creates a varying blur that synthetic training data cannot fully replicate. This often leads existing methods to either smooth out real details or invent fake astronomical features. FluxFlow proposes a flow-matching model trained in pixel space that factors in measurement uncertainties and prioritizes important source regions, then applies a Wiener filter correction during testing to keep outputs physically plausible. A new benchmark of nearly 20,000 real paired images from DESI ground and HST space telescopes shows the approach beats prior techniques on both brightness accuracy and scientific reliability.

Core claim

FluxFlow establishes that conservative pixel-space flow-matching, when trained with observation uncertainty and source-region importance weights and paired with a training-free Wiener-regularized correction at test time, delivers super-resolved astronomical images with superior photometric and scientific accuracy on real ground-to-space pairs compared to baseline methods.

What carries the argument

Conservative pixel-space flow-matching framework that integrates observation uncertainty, source-region importance weights during training, and a Wiener-regularized test-time correction.

Load-bearing premise

That weighting training by observation uncertainty and source importance, combined with the Wiener correction, will consistently prevent hallucinations across varying real atmospheric conditions not fully represented in the dataset.

What would settle it

If FluxFlow produces images with more hallucinated features or poorer scientific metrics than simpler methods when tested on fresh, independent ground-to-space image pairs with different seeing conditions, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2605.03749 by Dong Li, Gengjia Chang, Jun Liu, Lin Gu, Liuzhuozheng Li, Shuhong Liu, Tatsuya Harada, Xining Ge, Xuangeng Chu, Ziteng Cui, Ziying Gu.

Figure 1
Figure 1. Figure 1: FluxFlow reconstructs HST-quality high-resolution images from seeing-limited low view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of point source degradation from ideal (left) to diffraction-limited (center) to view at source ↗
Figure 3
Figure 3. Figure 3: Paired examples from our DESI–HST dataset. Top: DESI cutouts at 128 view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of HST sci￾ence images (top) and WHT weight maps (bottom), showing structures from the drizzle stacking and local artifacts such as cosmic rays. Inverse-Variance Weighting. The drizzled HST mosaic is accompanied by a per-pixel weight map (WHT) W ∈ R Hhr×Whr encoding the inverse variance at each pixel. It reflects spatially varying exposure depth from the dither pat￾tern together with statisti… view at source ↗
Figure 5
Figure 5. Figure 5: Sampling trajectory of a representative hallucination source across ODE integration steps. view at source ↗
Figure 6
Figure 6. Figure 6: Aperture flux (sum) evolution of bright hallucinated sources along the ODE sampling trajectory. It starts from large￾variance noise and converges to brightness levels exceeding the HST (red) at t=1. Hallucination Sources. We track hallucinations along the full ODE trajectory. As shown in view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of baseline methods and FluxFlow on view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison of baseline methods and FluxFlow on view at source ↗
Figure 9
Figure 9. Figure 9: WHT weighting ablation. Defects (cosmic rays, drizzle gaps) in red boxes. DESI w/o MCFS w/o Wiener Ours HST view at source ↗
Figure 10
Figure 10. Figure 10: MC-FS and Wiener-deconvolved cor￾rection ablation. Hallucinations in red boxes view at source ↗
Figure 11
Figure 11. Figure 11: The two instruments behind our data. Left: the Hubble Space Telescope (HST), the space view at source ↗
Figure 12
Figure 12. Figure 12: The footprint of the Legacy Surveys (DR10) i-band observations (dark red) overlaid with view at source ↗
Figure 13
Figure 13. Figure 13: Architecture of the FluxFlow velocity network view at source ↗
Figure 14
Figure 14. Figure 14: Conceptual illustration of MC-FS. The flow transports view at source ↗
Figure 15
Figure 15. Figure 15: Toy reconstruction at σn = 0 with N = 20 Euler steps. The four panels show, from left to right, the HST-quality truth, the DESI observation, the naive adjoint back-projection, and the Wiener-corrected update. The naive recovery has fitted width σfit ≈ 2.75, visibly broadened relative to the truth despite the observation being noise-free. The Wiener-corrected update at the same N has σfit ≈ 2.03, restored … view at source ↗
Figure 16
Figure 16. Figure 16: Numerical verification of the closed-form analysis at the same parameters as Figure view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative comparison of the latent-space flow-matching ablation. Four representative view at source ↗
Figure 18
Figure 18. Figure 18: Effect of the forward PSF width σPSF used inside MC-FS at ×2 (top) and ×4 (bottom). PSNR decreases monotonically with σPSF, while Flux-L1 and Detection F1 form broad plateaus over an intermediate range. We set σPSF = 2 HR pixels at both scales in our experiments. gaps than the latent-space variant but consistent in direction across PSNR, Flux-L1, and detection F1. The interpretation is that the bridge sta… view at source ↗
Figure 19
Figure 19. Figure 19: Effect of the Wiener regularization λSNR at ×2. Sweeping λSNR from 10 to 100 leaves PSNR, Flux-L1, and Detection F1 essentially unchanged, in agreement with the spectral analysis of Sections E.2 and E.3. Our default λSNR = 50 is the value used throughout view at source ↗
Figure 20
Figure 20. Figure 20: Effect of the base Wiener correction step size view at source ↗
Figure 21
Figure 21. Figure 21: MC-FS hallucination suppression on three DESI–HST examples. Columns show the view at source ↗
Figure 21
Figure 21. Figure 21: Effect of the number of sampling steps T at ×2, with all other MC-FS hyperparameters held at their defaults. PSNR decreases monotonically with T; Flux-L1 and Detection F1 attain their joint optimum at T = 10 before degrading at larger T as the per-step Wiener correction ηi shrinks faster than the cumulative ODE drift along vθ. Our default T = 10 is the value used throughout view at source ↗
Figure 22
Figure 22. Figure 22: Additional MC-FS detection comparisons under the protocol of Figure view at source ↗
Figure 22
Figure 22. Figure 22: MC-FS hallucination suppression on three DESI–HST examples. Columns show the view at source ↗
Figure 23
Figure 23. Figure 23: Qualitative DESI–HST ×2 comparison of all evaluated methods, with two zoom-in panels per method. FluxFlow recovers the diffraction spikes of the central star and the morphology of the extended source while avoiding the speckled background of AS-Bridge. DESI Bicubic SwinIR HAT FISR cGAN GD-NET AS-Bridge Ours HST view at source ↗
Figure 23
Figure 23. Figure 23: Additional MC-FS detection comparisons under the protocol of Figure view at source ↗
Figure 24
Figure 24. Figure 24: Additional DESI–HST ×2 comparison, on a bright central star surrounded by closely packed compact sources. FluxFlow reproduces the spike pattern and resolves the neighboring sources without the merging or smearing seen in the baselines. 31 view at source ↗
Figure 25
Figure 25. Figure 25: DESI–HST ×4 comparison on a field with a saturated source on the upper-left corner and several low-SNR background galaxies. FluxFlow preserves faint background galaxies that Bicubic and GD-Net erase while reconstructing the saturated halo more cleanly than the speckled output of AS-Bridge. DESI Bicubic SwinIR HAT FISR cGAN GD-NET AS-Bridge Ours HST view at source ↗
Figure 25
Figure 25. Figure 25: Additional DESI–HST ×2 comparison, on a bright central star surrounded by closely packed compact sources. FluxFlow reproduces the spike pattern and resolves the neighboring sources without the merging or smearing seen in the baselines. 31 view at source ↗
Figure 26
Figure 26. Figure 26: DESI–HST ×4 comparison on a binary-star field. FluxFlow reproduces the four-pointed diffraction pattern on the lower companion together with surrounding faint sources, whereas re￾gression baselines suppress the spike pattern and the generative baselines either oversmooth or oversharpen the background. 32 view at source ↗
Figure 27
Figure 27. Figure 27: DESI–HST ×4 comparison on a binary-star field. FluxFlow reproduces the four-pointed diffraction pattern on the lower companion together with surrounding faint sources, whereas re￾gression baselines suppress the spike pattern and the generative baselines either oversmooth or oversharpen the background. 32 view at source ↗
read the original abstract

Ground-to-space astronomical super-resolution requires recovering space-quality images from ground-based observations that are simultaneously limited by pixel sampling resolution and atmospheric seeing, which imposes a stochastic, spatially varying PSF that cannot be resolved through upsampling alone. Existing methods rely on synthetic training pairs that fail to capture real atmospheric statistics and are prone to either over-smoothed reconstructions or hallucination sources with no physical counterpart in the observed sky. We propose FluxFlow, a conservative pixel-space flow-matching framework that incorporates observation uncertainty and source-region importance weights during training, and a training-free Wiener-regularized test-time correction to suppress hallucination sources while preserving recovered detail. We further construct the DESI--HST Dataset, the large-scale real-world benchmark comprising 19,500 real co-registered ground-to-space image pairs with real atmospheric PSF variation. Experiments demonstrate that FluxFlow consistently outperforms existing baseline methods in both photometric and scientific accuracy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The manuscript introduces FluxFlow, a pixel-space flow-matching framework for ground-to-space astronomical image super-resolution. It incorporates observation uncertainty and source-region importance weights during training, along with a training-free Wiener-regularized test-time correction to mitigate hallucinations while preserving detail. The authors also construct the DESI-HST dataset of 19,500 real co-registered ground-to-space image pairs exhibiting real atmospheric PSF variation and report that FluxFlow outperforms existing baselines in both photometric and scientific accuracy.

Significance. If the experimental claims hold, the work could meaningfully advance astronomical imaging by shifting from synthetic training pairs (which fail to capture real atmospheric statistics) to real paired data and a conservative flow-matching approach explicitly designed to suppress non-physical hallucinations. The emphasis on uncertainty weighting, importance sampling, and the Wiener correction provides a principled way to handle spatially varying PSFs, which is load-bearing for scientific usability of the outputs.

minor comments (3)
  1. The abstract states that FluxFlow 'consistently outperforms existing baseline methods in both photometric and scientific accuracy' but provides no numerical values, error bars, or specific metrics (e.g., RMSE, PSNR, or scientific figure-of-merit). Adding one or two key quantitative results would strengthen the summary.
  2. Section describing the Wiener-regularized test-time correction should explicitly state the regularization parameter schedule and how it interacts with the flow-matching velocity field; the current description leaves the exact implementation of the 'conservative' property somewhat implicit.
  3. The DESI-HST dataset construction paragraph would benefit from a brief statement on the co-registration accuracy and the distribution of seeing FWHM values across the 19,500 pairs to allow readers to assess how representative the benchmark is of typical ground-based conditions.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our work on FluxFlow and the recommendation for minor revision. The referee summary correctly describes the pixel-space flow-matching framework, the use of uncertainty and importance weights, the Wiener test-time correction, and the construction of the real DESI-HST dataset. We are pleased that the potential impact on handling real atmospheric PSFs and suppressing non-physical hallucinations is recognized.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's method is presented as a pixel-space flow-matching framework that augments established flow-matching with observation uncertainty weighting, source-region importance, and a training-free Wiener-regularized correction; none of these elements are shown to be defined in terms of the target photometric or scientific accuracy metrics. The central performance claims rest on experiments using the newly introduced DESI-HST real co-registered pairs rather than on any self-referential fit or prediction that reduces to the inputs by construction. No load-bearing self-citation chains, uniqueness theorems, or ansatz smuggling are indicated in the provided description, leaving the derivation self-contained against external benchmarks and real data.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the method description references established flow-matching and Wiener filtering without introducing new postulated quantities.

pith-pipeline@v0.9.0 · 5487 in / 1186 out tokens · 52946 ms · 2026-05-08T18:51:36.505049+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

23 extracted references · 19 canonical work pages · 8 internal anchors

  1. [1]

    Beyond Model Design: Data-Centric Training and Self-Ensemble for Gaussian Color Image Denoising

    Gengjia Chang, Xining Ge, Weijun Yuan, Zhan Li, Qiurong Song, Luen Zhu, and Shuhong Liu. Beyond model design: Data-centric training and self-ensemble for gaussian color image denoising. arXiv preprint arXiv:2604.11468, 2026a. Gengjia Chang, Xining Ge, Weijun Yuan, Zhan Li, Qiurong Song, Luen Zhu, and Shuhong Liu. Training-free model ensemble for single-im...

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Diffusion Posterior Sampling for General Noisy Inverse Problems

    Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super-resolution transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367–22377, 2023a. Xiangyu Chen, Xintao Wang, Wenlong Zhang, Xiangtao Kong, Yu Qiao, Jiantao Zhou, and Chao Dong. Hat: Hybrid attention tr...

    work page internal anchor Pith review arXiv

  3. [3]

    Training Flow Matching: The Role of Weighting and Parameterization

    Anne Gagneux, Ségolène Martin, Rémi Gribonval, and Mathurin Massias. Training flow matching: The role of weighting and parameterization.arXiv preprint arXiv:2603.06454,

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Dual-Branch Remote Sensing Infrared Image Super-Resolution

    Xining Ge, Gengjia Chang, Weijun Yuan, Zhan Li, Zhanglu Chen, Boyang Yao, Yihang Chen, Yifan Deng, and Shuhong Liu. Dual-branch remote sensing infrared image super-resolution.arXiv preprint arXiv:2604.10112, 2026a. Xining Ge, Weijun Yuan, Gengjia Chang, Xuyang Li, and Shuhong Liu. Clip-guided data augmenta- tion for night-time image dehazing.arXiv preprin...

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    Classifier-Free Diffusion Guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598,

    work page internal anchor Pith review arXiv

  6. [6]

    Bbdm: Image-to-image translation with brownian bridge diffusion models

    Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. Bbdm: Image-to-image translation with brownian bridge diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1952–1961,

    1952

  7. [7]

    arXiv preprint arXiv:2502.01993 (2025)

    Boyun Li, Haiyu Zhao, Wenxin Wang, Peng Hu, Yuanbiao Gou, and Xi Peng. Mair: A locality-and continuity-preserving mamba for image restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7491–7501, 2025a. Jianze Li, Jiezhang Cao, Yong Guo, Wenbo Li, and Yulun Zhang. One diffusion step to real-world super-reso...

    work page arXiv

  8. [8]

    Back to Basics: Let Denoising Generative Models Denoise

    Tianhong Li and Kaiming He. Back to basics: Let denoising generative models denoise.arXiv preprint arXiv:2511.13720,

    work page internal anchor Pith review arXiv

  9. [9]

    A self-attention based srgan for super-resolution of astro- nomical image

    Wanjun Li, Zhe Liu, and Hongtao Deng. A self-attention based srgan for super-resolution of astro- nomical image. In2022 IEEE 8th International Conference on Computer and Communications, pages 1977–1981. IEEE,

    1977

  10. [10]

    Flow Matching for Generative Modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

    work page Pith review arXiv

  11. [11]

    I2SB: Image-to-image Schrödinger bridge.arXiv preprint arXiv:2302.05872,

    Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A Theodorou, Weili Nie, and Anima Anandkumar. I2SB: Image-to-image Schrödinger bridge.arXiv preprint arXiv:2302.05872,

    work page arXiv

  12. [12]

    Realx3d: A physically-degraded 3d benchmark for multi-view visual restoration and recon- struction.arXiv preprint arXiv:2512.23437, 2025

    Shuhong Liu, Chenyu Bao, Ziteng Cui, Yun Liu, Xuangeng Chu, Lin Gu, Marcos V Conde, Ryo Umagami, Tomohiro Hashimoto, Zijian Hu, et al. Realx3d: A physically-degraded 3d benchmark for multi-view visual restoration and reconstruction.arXiv preprint arXiv:2512.23437, 2025a. Shuhong Liu, Xiang Chen, Hongming Chen, Quanfeng Xu, and Mingrui Li. Deraings: Gaussi...

    work page arXiv 2026

  13. [13]

    Fine-grained image editing by pixel-wise guidance using diffusion models.arXiv preprint arXiv:2212.02024,

    Naoki Matsunaga, Masato Ishii, Akio Hayakawa, Kenji Suzuki, and Takuya Narihira. Fine-grained image editing by pixel-wise guidance using diffusion models.arXiv preprint arXiv:2212.02024,

    work page arXiv

  14. [14]

    Diamond: Directed inference for artifact mitigation in flow matching models

    Alicja Polowczyk, Agnieszka Polowczyk, Piotr Borycki, Joanna Waczy ´nska, Jacek Tabor, and Przemysław Spurek. Diamond: Directed inference for artifact mitigation in flow matching models. arXiv preprint arXiv:2602.00883,

    work page arXiv

  15. [15]

    Flower: A flow-matching solver for inverse problems.arXiv preprint arXiv:2509.26287,

    Mehrsa Pourya, Bassam El Rawas, and Michael Unser. Flower: A flow-matching solver for inverse problems.arXiv preprint arXiv:2509.26287,

    work page arXiv

  16. [16]

    Generative AI for image reconstruction in Intensity Interferometry: a first attempt

    Km Nitu Rai, Yuri van der Burg, Soumen Basak, Prasenjit Saha, and Subrata Sarangi. Generative ai for image reconstruction in intensity interferometry: a first attempt.arXiv preprint arXiv:2508.03398,

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    The eleventh NTIRE 2026 efficient super-resolution challenge report

    Bin Ren, Hang Guo, Yan Shu, Jiaqi Ma, Ziteng Cui, Shuhong Liu, Guofeng Mei, Lei Sun, Zongwei Wu, Fahad Shahbaz Khan, Salman Khan, Radu Timofte, Yawei Li, et al. The eleventh NTIRE 2026 efficient super-resolution challenge report. InProceedings of the Computer Vision and Pattern Recognition Conference Workshops,

    2026

  18. [18]

    Palette: Image-to-image diffusion models

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. InACM SIGGRAPH 2022 Conference Proceedings, pages 1–10, 2022a. Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative re...

    2022

  19. [19]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456,

    work page Pith review arXiv 2011

  20. [20]

    Improving and generalizing flow-based generative models with minibatch optimal transport

    Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector- Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport.arXiv preprint arXiv:2302.00482,

    work page internal anchor Pith review arXiv

  21. [21]

    Star: A benchmark for astronomical star fields super-resolution.arXiv preprint arXiv:2507.16385,

    Kuo-Cheng Wu, Guohang Zhuang, Jinyang Huang, Xiang Zhang, Wanli Ouyang, and Yan Lu. Star: A benchmark for astronomical star fields super-resolution.arXiv preprint arXiv:2507.16385,

    work page arXiv

  22. [22]

    Dichang Zhang, Yixuan Shao, Simon Birrer, and Dimitris Samaras

    URL https://openreview.net/forum?id=TwvnG6mobR. Dichang Zhang, Yixuan Shao, Simon Birrer, and Dimitris Samaras. As-bridge: A bidirectional gener- ative framework bridging next-generation astronomical surveys.arXiv preprint arXiv:2603.11928,

    work page arXiv

  23. [23]

    Guided flows for generative modeling and decision making.arXiv preprint arXiv:2311.13443, 2023

    Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient image super-resolution. InEuropean Conference on Computer Vision, pages 483–500. Springer, 2024a. Yasi Zhang, Peiyu Yu, Yaxuan Zhu, Yingshan Chang, Feng Gao, Ying N Wu, and Oscar Leong. Flow priors for linear inverse problems via iterative corrupted trajectory matching...

    work page arXiv