FluxFlow: Conservative Flow-Matching for Astronomical Image Super-Resolution
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "FluxFlow: Conservative Flow-Matching for Astronomical Image Super-Resolution".
Jane: Here's the gist of the paper. 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 <ref:2605.03749#pg2> This paper proposes FluxFlow, a conservative pixel-space flow-matching framework that incorporates observation uncertainty and source-region importance weights during training, and a training-free Wienerregularized test-time correction to suppress hallucination sources while preserving recovered detail <ref:2605.03749#pg3> Experiments demonstrate that FluxFlow consistently outperforms existing baseline methods in both photometric and scientific accuracy <ref:2605.03749#pg2> Pruning redundant reasoning tokens The core problem addressed is the gap between wide-field ground-based surveys and high-resolution space-based observations, where atmospheric turbulence broadens the Point Spread Function (PSF) to a seeing-limited state <ref:2605.03749#pg3> Existing methods fail because they rely on synthetic training pairs that do not capture real atmospheric statistics and are prone to either over-smoothed reconstructions or hallucination sources with no physical counterpart in the observed sky <ref:2605.03749#pg2> The authors construct the DESI–HST Dataset, a large-scale real-world benchmark comprising 19,500 real co-registered ground-to-space image pairs with real atmospheric PSF variation <ref:2605.03749#pg2> How the training signal is built The FluxFlow framework utilizes a pixel-space flow matching approach where the velocity field is trained to minimize the Conditional Flow Matching (CFM) objective <ref:2605.03749#pg2> This objective, denoted as LCFM, regresses against the conditional velocity ut rather than the intractable marginal velocity field <ref:2605.03749#pg2> To incorporate physical fidelity, the loss is weighted by a spatially varying weight map LWFM = Et, x0, x1
P p Wf ⊙ S: vθ(xt, t) − ut2 p P p Wf ⊙ S <ref:2605.03749#pg2> This spatial weighting incorporates two factors: the Inverse-Variance Weighting (WHT) which reflects spatially varying exposure depth and noise, and the Source-Region Weighting (S) which flags scientifically important source pixels <ref:2605.03749#pg2> Where the gains come from The method introduces two key innovations for inference: a training-free Wienerregularized test-time correction to suppress hallucination sources and a measurement-consistent flow sampling scheme <ref:2605.03749#pg2> The measurement-consistent flow sampling enforces physical consistency by back-projecting residuals through a Wienerregularized inverse of the forward operator <ref:2605.03749#pg2> This correction uses a Wiener correction kernel W(k) derived from the Minimum Mean Square Error (MMSE) principle to replace the naive adjoint projection, preventing cumulative PSF widening across integration steps <ref:2605.03749#pg2> Evaluation and Scientific Accuracy FluxFlow achieves superior photometric fidelity and scientific accuracy over both recent regression-based and generative approaches on the DESI–HST dataset <ref:2605.03749#pg2> Quantitative results show that FluxFlow obtains PSNR competitive with regression baselines, while achieving the best SSIM and Flux-L1 among all compared methods <ref:2605.03749#pg2> Downstream scientific applications confirm this superiority, as FluxFlow achieves the best mean and the best median on four of the six metrics on a matched bright-source set, dominating magnitude error and luminosity-proxy stellar-mass MAE <ref:2605.03749#pg2> The suppression of spurious sources in low-SNR background regions demonstrates the benefit of the MC-FS correction over naive OT-CFM sampling <ref:2605.03749#pg2> Limitations A limitation noted is that hallucinations are suppressed rather than eliminated, and scientific accuracy still falls short of what precision astronomical measurements demand <ref:2605.03749#pg2> Future work will extend FluxFlow to multi-band joint reconstruction and uncertainty-calibrated sampling to improve the SNR <ref:2605.03749#pg2> The analysis shows that the dominant pathology at practical finite N is the low-pass effect of Equation (27) rather than the asymptotic noise amplification of Equation (30) <ref:2605.03749#pg2> In summary, FluxFlow recovers fine details including diffraction spikes around bright stars and sharp cores of point sources while preserving the shape and orientation of extended galaxies <ref:2605.03749#pg2> The method's robustness is confirmed by its stability across variations in the forward PSF width σPSF and Wiener regularization parameter λSNR <ref:2605.03749#pg2> The final results show that FluxFlow produces reconstructions that combine perceptual sharpness with morphological and photometric fidelity, as required for scientific use of astronomical imagery <ref:2605.03749#pg2> The paper concludes by presenting FluxFlow, a conservative flow-matching framework for ground-to-space astronomical super-resolution <ref:2605.03749#pg2> The authors report that the full model attains the strongest photometric scores together with the best detection F1, confirming that both inference-time components are necessary <ref:2605.03749#pg2> The system is robust to mis-specification of λSNR by an order of magnitude in either direction, and its default settings are well within a stable plateau <ref:2605.03749#pg2> The overall performance indicates that the inverse-variance and source-region weighting together with the Wiener-corrected MC-FS sampling translate directly into the per-source flux fidelity that downstream catalog products require <ref:2605.03749#pg2> The final comparison across all four test cases shows FluxFlow recovers diffraction patterns, faint background structures, and source morphology that regression baselines oversmooth and that generative baselines either speckle or oversharpen <ref:2605.03749#pg2> The paper is supported by a comprehensive set of references covering flow matching, diffusion models, and astronomical image restoration techniques <ref:2605.03749#pg2> Page 1 --- FluxFlow: Conservative Flow-Matching for Astronomical Image Super-Resolution Shuhong Liu1,2, Xining Ge2,, Ziteng Cui1,2, Liuzhuozheng Li1 Gengjia Chang2, Jun Liu2, Ziying Gu1, Dong Li2, Xuangeng Chu1,2 Lin Gu3, Tatsuya Harada1,4 1The University of Tokyo 2 I2WM 3Tohoku University 4RIKEN AIP 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 <ref:2605.03749#pg2> 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 <ref:2605.03749#pg2> 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 Wienerregularized test-time correction to suppress hallucination sources while preserving recovered detail <ref:2605.03749#pg2> Pruning redundant reasoning tokens 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 <ref:2605.03749#pg2> 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 <ref:2605.03749#pg2> How the training signal is built During training, it uses spatial reliability and source-importance maps to focus learning on scientifically focused regions while suppressing unstable gradients from artifacts <ref:2605.03749#pg2> The overall objective is defined as LWFM = Et, x0, x1
P p Wf ⊙ S: vθ(xt, t) − ut2 p P p Wf ⊙ S <ref:2605.03749#pg2> Where the gains come from During inference, it enforces measurement consistency with the observed low-resolution image by back-projecting residuals through a Wiener-regularized inverse of the forward operator to suppress hallucination sources while preserving recovered detail <ref:2605.03749#pg2> This correction is implemented using Equation (11) which replaces the PSF adjoint in A⊤ with the Wiener kernel W(k) <ref:2605.03749#pg2> Evaluation and Scientific Accuracy Experiments on our curated dataset demonstrate that FluxFlow achieves superior photometric fidelity and scientific accuracy over both recent regression-based and generative approaches <ref:2605.03749#pg2> Table 2 shows quantitative results at ×2 and ×4 on our DESI–HST super-resolution dataset, where FluxFlow obtains PSNR competitive with regression baselines, within 0.1 dB of the highest value at both scales, while achieving the best SSIM and Flux-L1 among all compared methods <ref:2605.03749#pg2> The results confirm that FluxFlow recovers fine details including the diffraction spikes around bright stars and the sharp cores of point sources <ref:2605.03749#pg2> Limitations Hallucinations are suppressed rather than eliminated, and scientific accuracy still falls short of what precision astronomical measurements demand <ref:2605.03749#pg2> Future work will extend FluxFlow to multi-band joint reconstruction and uncertainty-calibrated sampling to improve the SNR <ref:2605.03749#pg2> The paper is supported by a comprehensive set of references covering flow matching, diffusion models, and astronomical image restoration techniques <ref:2605.03749#pg2> The final comparison across all four test cases shows FluxFlow recovers diffraction patterns, faint background structures, and source morphology that regression baselines oversmooth and that generative baselines either speckle or oversharpen <ref:2605.03749#pg2> The paper is supported by a comprehensive set of references covering flow matching, diffusion models, and astronomical image restoration techniques <ref:2605.03749#pg2> The final comparison across all four test cases shows FluxFlow recovers diffraction patterns, faint background structures, and source morphology that regression baselines oversmooth and that generative baselines either speckle or oversharpen <ref:2605.03749#pg2> Page 2 --- Ideal Source Diffraction-limited Airy Disk Seeing-limited Observation Figure 2 Illustration of point source degradation from ideal (left) to diffraction-limited (center) to seeing-limited (right), with 2D intensity images and 1D cross-sectional profiles <ref:2605.03749#pg2> This distinction is especially important in ground-to-space image super-resolution <ref:2605.03749#pg2> As illustrated in Figure 2, Ground-based observations are typically seeing-limited, where atmospheric turbulence broadens the PSF and imposes an angular-resolution floor that often dominates over the detector sampling scale <ref:2605.03749#pg2> In contrast, space-based telescopes such as the Hubble Space Telescope (HST) are free from atmospheric seeing and can approach the diffraction limit of their optics, yielding substantially sharper images <ref:2605.03749#pg2> Bridging this gap would enable high-resolution analysis over survey scales that are inaccessible to space telescopes alone <ref:2605.03749#pg2> Pruning redundant reasoning tokens Ideal Source Diffraction-limited Airy Disk Seeing-limited Observation Figure 2 Illustration of point source degradation from ideal (left) to diffraction-limited (center) to seeing-limited (right), with 2D intensity images and 1D cross-sectional profiles <ref:2605.03749#pg2> This distinction is especially important in ground-to-space image super-resolution <ref:2605.03749#pg2> As illustrated in Figure 2, Ground-based observations are typically seeing-limited, where atmospheric turbulence broadens the PSF and imposes an angular-resolution floor that often dominates over the detector sampling scale <ref:2605.03749#pg2> In contrast, space-based telescopes such as the Hubble Space Telescope (HST) are free from atmospheric seeing and can approach the diffraction limit of their optics, yielding substantially sharper images <ref:2605.03749#pg2> Bridging this gap would enable high-resolution analysis over survey scales that are inaccessible to space telescopes alone <ref:2605.03749#pg2> How the training signal is built The authors construct the DESI–HST Dataset, a real paired ground-to-space super-resolution benchmark comprising 19,500 real co-registered ground-to-space image pairs with real atmospheric PSF variation and observation uncertainty maps <ref:2605.03749#pg2> Where the gains come from The training objective uses spatial reliability and source-importance maps to focus learning on scientifically focused regions while suppressing unstable gradients from artifacts <ref:2605.03749#pg2> The inference correction enforces measurement consistency with the observed low-resolution image by back-projecting residuals through a Wiener-regularized inverse of the forward operator <ref:2605.03749#pg2> Evaluation and Scientific Accuracy Experiments on our curated dataset demonstrate that FluxFlow achieves superior photometric fidelity and scientific accuracy over both recent regression-based and generative approaches <ref:2605.03749#pg2> Table 2 shows quantitative results at ×2 and ×4 on our DESI–HST super-resolution dataset, where FluxFlow obtains PSNR competitive with regression baselines, within 0.1 dB of the highest value at both scales, while achieving the best SSIM and Flux-L1 among all compared methods <ref:2605.03749#pg2> Limitations The scarcity of real observational data for ground-to-space SR is overcome by constructing the DESI–HST Dataset, the large-scale astronomical SR dataset comprising real-world co-registered ground-to-space image pairs with authentic atmospheric PSF variation and observation uncertainty maps <ref:2605.03749#pg2> Page 3 --- 滚滚长江东逝水 Figure 3 Paired examples from our DESI–HST dataset.
Tom: Top: DESI cutouts at 128×128 pixels; Bottom: HST observations at 512×512 pixels.
Jane: All pairs are pixel-aligned <ref:2605.03749#pg2> Pruning redundant reasoning tokens Figure 3 Paired examples from our DESI–HST dataset.
Tom: Top: DESI cutouts at 128×128 pixels; Bottom: HST observations at 512×512 pixels.
Title and authors: Tom: So we're talking about this new paper called FluxFlow: Conservative Flow-Matching for Astronomical Image Super-Resolution. Jane, what's the core problem they're trying to solve with this research?
Jane: Well, Tom, the main thing is bridging that gap between ground-based surveys and space telescope images. Ground observations are limited by atmospheric seeing, which makes the point spread function change randomly across the image—it’s spatially varying.
Lu: Exactly. That spatial variation in seeing creates a kind of messy blur that simple upsampling just can't fix because it doesn't understand that the blur is changing depending on where you are looking <ref:2605.03749#pg2>.
Meng: And existing methods, they rely on synthetic training pairs, which isn't real atmospheric statistics, so they either over-smooth things or they hallucinate stuff that isn't actually there in the sky <ref:2605.03749#pg2>.
Tom: So FluxFlow is proposing a pixel-space flow matching framework that tries to learn this messy process while incorporating observation uncertainty and source importance during training, and then it has this trick for test time to stop those hallucinations without losing the detail <ref:2605.03749#pg1>.
Jane: It's smart because they are building a real-world benchmark called the DESI–HST Dataset, which has nineteen thousand five hundred real pairs of ground and space images with actual atmospheric variations <ref:2605.03749#pg2>.
Lu: During training, they focus the learning on scientifically important regions by using spatial reliability and source-importance maps to suppress unstable gradients from artifacts <ref:2605.03749#pg2>.
Tom: And the objective function they use is this thing, LWFM equals E t x zero x one P p W f sigma S v theta at t <ref:2605.03749#pg1>. It’s basically a weighted loss that uses exposure depth and noise as weights, plus a source mask to flag important pixels.
Jane: That weighting is key because it incorporates both the inverse variance of the pixel, which handles noise differently at different spots, and the source region importance map <ref:2605.03749#pg2>.
Lu: The inference part has two big innovations here—a training-free Wiener regularized test time correction to stop hallucinations, and a measurement consistent flow sampling scheme <ref:2605.03749#pg1>.
Meng: How does that test time correction actually work? Because we’re shipping models, I need to know what breaks when it runs on real traffic.
Title and authors: Tom: It enforces consistency with the observation by back-projecting residuals through a Wiener regularized inverse of the forward operator <ref:2605.03749#pg1>. They use a Wiener correction kernel, W k, which comes from the minimum mean square error principle to replace the naive adjoint projection <ref:2605.03749#pg1>.
Jane: That prevents that cumulative widening of the point spread function you see when you integrate through multiple steps <ref:2605.03749#pg2>.
Lu: And they use this consistency enforcement to suppress spurious sources in low signal regions, which is a big win compared to just using naive adjoint projection <ref:2605.03749#pg1>.
Tom: The evaluation shows FluxFlow achieves competitive PSNR with regression baselines, within zero point one decibel of the highest value at both scale two and four on their DESI–HST dataset <ref:2605.03749#pg2>.
Jane: But they also claim it gets the best SSIM and Flux-L1 among all compared methods, which suggests it’s not just about pixel matching, but about actual visual fidelity <ref:2605.03749#pg2>.
Lu: The results confirm that this method recovers fine details like diffraction spikes around bright stars and the sharp cores of point sources, while also preserving the shape and orientation of extended galaxies <ref:2605.03749#pg1>.
Meng: If I were deploying this, what's the runtime like? Does this flow matching thing add a huge computational cost compared to just running a standard CNN restoration model?
Tom: It’s conservative, so it’s designed to be practical, and the stability across variations in the forward point spread function width sigma PSF and Wiener regularization parameter lambda SNR is quite good <ref:2605.03749#pg2>.
Jane: So if you're just looking at what this means for us on the ground—for things like galaxy surveys—it suggests we can get space-quality resolution over survey scales that were previously inaccessible to space telescopes alone <ref:2605.03749#pg2>.
Lu: That bridging of the gap is the big idea here, because as you can see in Figure two ground observations are seeing limited, while HST is diffraction limited <ref:2605.03749#pg2>.
Tom: So to wrap up on FluxFlow: it’s a conservative flow matching framework that uses observation uncertainty and source importance weights during training, and then adds a training-free Wiener regularized test time correction to suppress hallucination while keeping detail, achieving the best photometric scores over regression methods <ref:2605.03749#pg1>.
Title and authors: Jane: It seems like they've really figured out how to make the math work in a way that respects the real physics of atmospheric blurring and measurement noise.
Lu: They also noted a limitation: hallucinations are suppressed rather than completely eliminated, and scientific accuracy still doesn't meet what precision astronomical measurements demand <ref:2605.03749#pg2>.
Meng: That makes sense for now. If I were to push this further, I'd want to see how they handle multi-band joint reconstruction next, because that would add another layer of complexity to the deployment pipeline <ref:2605.03749#pg2>.
Tom: Yeah, extending it for multi-band will definitely test the robustness of those weights and the correction scheme we talked about <ref:2605.03749#pg2>.
Jane: So, to wrap up on FluxFlow: it’s a conservative flow matching framework for astronomical image super-resolution that uses observation uncertainty and source importance weights during training, and then adds a training-free Wiener regularized test time correction to suppress hallucination while keeping detail <ref:2605.03749#pg1>.
Lu: It’s a method that successfully combines perceptual sharpness with morphological fidelity, which is what you need for actual scientific use of astronomical imagery <ref:2605.03749#pg1>.
Meng: And the fact that it's stable across variations in the forward point spread function width sigma PSF and Wiener regularization parameter lambda SNR shows it’s reasonably robust when you push those settings <ref:2605.03749#pg2>.
Tom: It seems like they've shown that the inverse variance and source-region weighting together with the Wiener corrected flow sampling translates directly into the per-source flux fidelity that downstream catalog products require <ref:2605.03749#pg2>.
Jane: That’s a big deal because it means we can trust these reconstructions for things like measuring luminosity proxies and stellar mass error, which is what they confirmed with their matched bright-source set <ref:2605.03749#pg2>.
Lu: Overall, FluxFlow recovers diffraction patterns, faint background structures, and source morphology that regression baselines oversmooth and generative baselines either speckle or oversharpen <ref:2605.03749#pg2>.
Tom: It’s a strong result because they've shown it can recover those fine details that other methods miss, even when you compare it against the best regression approaches <ref:2605.03749#pg2>.
Jane: So that’s what we have on FluxFlow for now—a method that balances the need for sharp images with keeping them physically consistent with what our instruments actually saw.
The paper's summary: Tom: So, to recap, FluxFlow is this new flow matching method for super-resolution that focuses on training using spatial reliability and source importance maps, and then it uses a special test time correction to stop those fake images from appearing.
Jane: Right, so it’s about making the model focus only on the parts of the image that are actually important scientifically and avoiding those messy artifacts we talked about before.
Lu: It’s reaching for something where old methods used synthetic data, and this one is using real pairings—the DESI-HST dataset—which has actual atmospheric variations, which is a big step up <ref:2605.03749#pg2>.
Meng: I'm interested in the training part because if you can build a loss function that respects both the noise level and where the light actually comes from, it makes sense for production.
Tom: It does, Lu. The objective function they use weights against both the inverse variance of each pixel and a source mask <ref:2605.03749#pg1>. It’s not just one loss anymore; it’s factoring in the physics of the observation directly into how the model learns to map pixels together.
Jane: That sounds like they are telling the AI, "Hey, learn this relationship better where your data is clean and scientifically valuable."
Lu: Exactly. And then for inference, they aren't just doing a regular projection; they're back-projecting the residuals through a Wiener regularized inverse of the forward operator <ref:2605.03749#pg1>. It uses this Wiener correction kernel to smooth out those errors as you go, which stops the point spread function from getting wider too much over multiple steps.
Meng: So, it keeps things stable when you run the inference pipeline, which is what matters when we try to deploy these things on real data. I’m just worried about the cost of running that Wiener correction kernel during inference.
Tom: The paper shows it's robust and stable across variations in the forward point spread function width sigma PSF and even changes in that Wiener regularization parameter lambda SNR <ref:2605.03749#pg2>. It’s behaving predictably, which is good for production deployment.
Jane: And what this means for us listening right now is that we can expect reconstructions of astronomical images—like those from ground-based telescopes—to be much more faithful to the real physics than what current regression or generative models are producing.
Lu: That’s because they showed it gets PSNR scores competitive with regression baselines, but simultaneously achieves the best SSIM and Flux-L1 metric <ref:2605.03749#pg2>. It recovers those fine details, like diffraction spikes around bright stars and the sharp cores of point sources, better than anything else compared on that dataset.
Meng: Competitive PSNR is one thing, but SSIM and Flux-L1 are what actually matter for visual fidelity. If you can get those metrics while keeping the noise suppression active during inference, that’s a solid result for deployment <ref:2605.03749#pg2>.
Tom: It's a strong claim because they show both training and test time components are necessary; you need both the smart way to learn and the smart way to correct when you actually use it.
Jane: So, while they haven't eliminated hallucinations—they’re suppressed—the overall picture is that FluxFlow recovers diffraction patterns and faint background structures in a way that regression methods oversmooth, and generative methods either speckle or oversharpen <ref:2605.03749#pg2>.
Lu: That ability to recover those specific features while maintaining photometric fidelity means this method translates directly into the per-source flux fidelity that catalog products actually require <ref:2605.03749#pg2>.
Meng: I'm just curious what they flag as a limitation, since we know perfect accuracy is hard to achieve in science.
Jane: They say that while hallucinations are suppressed, scientific accuracy still falls short of what precision astronomical measurements demand <ref:2605.03749#pg2>. They also point out that future work needs to look at multi-band joint reconstruction and uncertainty-calibrated sampling to improve the signal-to-noise ratio <ref:2605.03749#pg2>.
Tom: It seems like the next big thing for FluxFlow is taking it beyond just single bands, trying to handle more complex real-world data structures.
The paper's improvements: Tom: So, we're talking about how they suggest improving FluxFlow after they did the main work—they aren't just happy with their result, you know?
Jane: Right, so what’s next for this model? What are the authors saying needs to be done to make it even better for real science?
Lu: They’re looking at extending FluxFlow to multi-band joint reconstruction. That means they want the model to handle images taken with different colors or filters all at once <ref:2605.03749#pg2>.
Meng: Multi-band is a big deal for deployment, because real telescopes take data across many different wavelengths, so being able to reconstruct that holistically is crucial for a catalog product.
Tom: I think that’s smart because it forces the model to learn cross-channel relationships, not just one image at a time <ref:2605.03749#pg2>.
Jane: And they also talk about uncertainty-calibrated sampling. That means making sure the sampling itself is aware of how much noise there is in different areas, which should help improve that signal to noise ratio <ref:2605.03749#pg2>.
Lu: Yes, and they are trying to get better at handling those noise maps directly into the sampling process so the model doesn't just guess where the high-resolution detail should be <ref:2605.03749#pg2>.
Meng: From a production standpoint, if you can integrate uncertainty calibration into your sampling loop, it means you can build more reliable confidence scores for your reconstructed images on the fly. That's something I need to see in the next version <ref:2605.03749#pg2>.
Tom: It sounds like they’re moving from just image restoration to a full uncertainty-aware reconstruction pipeline, which is where you get closer to true scientific measurement capabilities.
Jane: So, it's about taking the model from just "making a pretty picture" to "making a scientifically trustworthy measurement," which is what we want in astronomy.
Lu: It really pushes the idea that these flow matching approaches can be built into systems that are inherently aware of observational noise and scientific importance, not just tacked on at the end <ref:2605.03749#pg1>.
Conclusion: Tom: So, to wrap up, FluxFlow is this conservative flow matching framework for astronomical image super resolution that uses observation uncertainty and source importance weights during training, and then adds a training free wiener regularized test time correction to suppress hallucination while preserving recovered detail.
Jane: It’s a method that successfully combines perceptual sharpness with morphological fidelity, which is what you need for actual scientific use of astronomical imagery.
Lu: And the fact that it's stable across variations in the forward point spread function width sigma PSF and wiener regularization parameter lambda SNR shows it’s reasonably robust when you push those settings <ref:2605.03749#pg2>.
Meng: It seems like they've shown that the inverse variance and source-region weighting together with the wiener corrected flow sampling translates directly into the per-source flux fidelity that downstream catalog products require <ref:2605.03749#pg2>.
Tom: And it confirms that both inference time components are necessary for getting those best photometric scores <ref:2605.03749#pg2>.
Jane: Overall, this paper shows we can bridge the gap between ground-based seeing and space-quality data by learning from real atmospheric physics rather than just synthetic pairs.
Lu: It’s a solid foundation for how flow matching can be made physically informed, which is a big step forward in representation learning for vision tasks <ref:2605.03749#pg1>.
Meng: I still see the limitation they mentioned: hallucinations are suppressed rather than eliminated, and scientific accuracy still falls short of what precision astronomical measurements demand <ref:2605.03749#pg2>.
Tom: True, but given the results on the DESI-HST dataset where they beat both regression and generative methods in SSIM and flux l one, it’s a very strong result for restoration work.
Jane: It means we can trust these reconstructions for things like measuring luminosity proxies and stellar mass error on survey scales that were previously inaccessible <ref:2605.03749#pg2>.
Lu: I think the next logical step is to extend FluxFlow to multi-band joint reconstruction, which would test the robustness of those weights across different spectral domains <ref:2605.03749#pg2>.
The University of Tokyo Institute for Advanced Study Tohoku University RIKEN AIP
cs.CV
Submitted: 2026-05-05
Updated: 2026-11-07
Comments: Accepted at the 40th Conference on Neural Information Processing Systems (NeurIPS 2026); to appear in Advances in Neural Information Processing Systems 39
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 100/100
The gist: 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…
Key concepts
- Pixel-Space Flow Matching
- This is the core learning approach where the model learns a velocity field that maps low-resolution pixels to high-resolution ones. Instead of predicting the entire complex flow at once, it trains on simpler conditional flows, making the training process more stable and focused on recovering fine details from ground observations.
- Source-Region Importance Weights (S)
- These weights identify which pixels in the image are scientifically important. By weighting the loss function based on these importance maps, the model learns to prioritize reconstructing features of significant sources, such as bright stars or galaxies, while ignoring noisy or less informative background regions.
- Wiener Regularized Test-Time Correction
- This is a crucial step during inference that corrects errors caused by atmospheric blurring. It uses a mathematical correction derived from the Minimum Mean Square Error principle to suppress 'hallucination' sources—false details created by the model—while ensuring the final image remains physically consistent with the original observation.
Terminology
Summary
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 <ref:2605. p. 0 FluxFlow: Conservative Flow-Matching for
Astronomical Image Super-Resolution
Shuhong Liu1,2, $
, Xining Ge2,
, Ziteng Cui1,2
, Liuzhuozheng Li1
Gengjia Chang2
, Jun Liu2
, Ziying Gu1
, Dong Li2
, Xuangeng Chu1,2
Lin Gu3
, Tatsuya Harada1,4
1The University of Tokyo 2
I2WM 3Tohoku University 4RIKEN AIP
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. 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 Wienerregularized 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.Where the paper says it
Where the paper says it
Where the paper says it
Pruning redundant reasoning tokens
The core problem addressed is the gap between wide-field ground-based surveys and high-resolution space-based observations, where atmospheric turbulence broadens the Point Spread Function (PSF) to a seeing-limited state <ref:2605.03749#pg3> Existing methods fail because they rely on synthetic training pairs that do not capture real atmospheric statistics and are prone to either over-smoothed reconstructions or hallucination sources with no physical counterpart in the observed sky <ref:2605. p. 0 FluxFlow: Conservative Flow-Matching for
Astronomical Image Super-Resolution
Shuhong Liu1,2, $
, Xining Ge2,
, Ziteng Cui1,2
, Liuzhuozheng Li1
Gengjia Chang2
, Jun Liu2
, Ziying Gu1
, Dong Li2
, Xuangeng Chu1,2
Lin Gu3
, Tatsuya Harada1,4
1The University of Tokyo 2
I2WM 3Tohoku University 4RIKEN AIP
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 Wienerregularized 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.Where the paper says it
Where the paper says it
How the training signal is built
The FluxFlow framework utilizes a pixel-space flow matching approach where the velocity field is trained to minimize the Conditional Flow Matching (CFM) objective <ref:2605.03749#pg2> This objective, denoted as LCFM, regresses against the conditional velocity ut rather than the intractable marginal velocity field <ref:2605. p. 3 [2023]:
LCFM = Et∼U[0,1], x0∼p0, x1∼p1
∥vθ(xt, t) − ut(xt | x0, x1)∥
2
, (4)
where xt = t x1 + (1 − t) x0 is the linearly interpolated sample. The CFM loss regresses against
the conditional velocity ut rather than the intractable marginal velocity field, yet provably yields the
same gradient in expectation Lipman et al. Overall Objective. Multiplying the two maps yields a single spatially varying weight that encodes
both reliability and importance, and is applied as a per-image normalized objective:
LWFM = Et, x0, x1
P
p
Wf ⊙ S
p
vθ(xt, t, c) − ut
2
p P
p
Wf ⊙ S
p
, (5)
where c denotes the conditioning upsampled DESI observation specified in Appendix D, ⊙ is
element-wise multiplication, p indexes spatial pixel locations, and the denominator P
p
(Wf ⊙ S)p is
recomputed per training pair so that each pair contributes a gradient signal of comparable magnitude
irrespective of the absolute scale of Wf ⊙ S. 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 pattern together with statistical noise and residual instrumental
artifacts such as chip gaps and cosmic-ray trails, as shown in
Figure 4.Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
The method introduces two key innovations for inference: a training-free Wienerregularized test-time correction to suppress hallucination sources and a measurement-consistent flow sampling scheme <ref:2605.03749#pg2> The measurement-consistent flow sampling enforces physical consistency by back-projecting residuals through a Wienerregularized inverse of the forward operator <ref:2605. p. 1 During training, it uses spatial reliability and source-importance maps to focus learning on
scientifically focused regions while suppressing unstable gradients from artifacts. During inference,
it enforces measurement consistency with the observed low-resolution image by back-projecting
residuals through a Wiener-regularized inverse of the forward operator. applying Equation (9) in ODE sampling would accumulate PSF-scale smoothing across integration
steps. This is because the adjoint A⊤ = H⊤
PSF ◦ D⊤
s
contains the same PSF convolution as the
forward operator A, so every correction step convolves the residual with the PSF kernel before adding
it back to x
(i+1).Where the paper says it
Where the paper says it
Evaluation and Scientific Accuracy
FluxFlow achieves superior photometric fidelity and scientific accuracy over both recent regression-based and generative approaches on the DESI–HST dataset <ref:2605.03749#pg2> Quantitative results show that FluxFlow obtains PSNR competitive with regression baselines, while achieving the best SSIM and Flux-L1 among all compared methods <ref:2605. p. 7 FluxFlow obtains
PSNR competitive with regression baselines, within 0. 1 dB of the highest value at both scales, while
achieving the best SSIM and Flux-L1 among all compared methods. In OT-CFM the velocity target ut = x1 − x0 is
independent of t and the interpolant xt = (1 − t) x0 + t x1 is a pixelwise convex combination, so
the mask M computed on x1 remains spatially aligned with source content at every t. 2 Measurement-Consistent Flow Sampling
The physics-informed weights improve velocity-field fidelity on source regions but leave the background unregulated.Where the paper says it
Where the paper says it
Limitations
A limitation noted is that hallucinations are suppressed rather than eliminated, and scientific accuracy still falls short of what precision astronomical measurements demand <ref:2605.03749#pg2> Future work will extend FluxFlow to multi-band joint reconstruction and uncertainty-calibrated sampling to improve the SNR <ref:2605. p. 8 Hallucinations are suppressed rather than eliminated, and scientific accuracy still falls
short of what precision astronomical measurements demand. Future work will extend FluxFlow to
multi-band joint reconstruction and uncertainty-calibrated sampling to improve the SNR. Hallucinations are suppressed rather than eliminated, and scientific accuracy still falls
short of what precision astronomical measurements demand. Future work will extend FluxFlow to
multi-band joint reconstruction and uncertainty-calibrated sampling to improve the SNR. Figure 7 presents qualitative comparisons at ×2. Regression-based methods
tend to oversmooth high-frequency structures, whereas FluxFlow recovers fine details including
the diffraction spikes around bright stars and the sharp cores of point sources. shape and orientation of extended galaxies and suppresses hallucinated sources in low signal-tonoise regions. FluxFlow thus produces reconstructions that combine perceptual sharpness with
morphological and photometric fidelity, as required for scientific use of astronomical imagery. Additional ablation studies on latent-space and bridge-based OT-CFM, hyperparameters,
and visualizations of detection-outcome are provided in Appendix G and Appendix I. 8 Conclusion
We present FluxFlow, a conservative flow-matching framework for ground-to-space astronomical
super-resolution. The broadened response smears
source flux beyond the target apertures into the background, and Table 3 reports a uniform photometric
degradation across PSNR, SSIM, and Flux-L1. Our full model attains the strongest photometric
scores together with the best detection F1, confirming that both inference-time components are
necessary. 8 Conclusion
We present FluxFlow, a conservative flow-matching framework for ground-to-space astronomical
super-resolution. The training objective weights the velocity loss by per-pixel inverse variance and
source-region importance, and the measurement-consistent sampling enforces physical consistency
through a Wiener-regularized back-projection. Regression-based methods
tend to oversmooth high-frequency structures, whereas FluxFlow recovers fine details including
the diffraction spikes around bright stars and the sharp cores of point sources. Generative methods
produce sharper textures but often introduce unreliable detail, whereas our model preserves the
8 Regression-based methods
tend to oversmooth high-frequency structures, whereas FluxFlow recovers fine details including
the diffraction spikes around bright stars and the sharp cores of point sources. Generative methods
produce sharper textures but often introduce unreliable detail, whereas our model preserves the
8 Regression-based methods
tend to oversmooth high-frequency structures, whereas FluxFlow recovers fine details including
the diffraction spikes around bright stars and the sharp cores of point sources. Generative methods
produce sharper textures but often introduce unreliable detail, whereas our model preserves the
8 FluxFlow: Conservative Flow-Matching for
Astronomical Image Super-Resolution
Shuhong Liu1,2, $
, Xining Ge2,
, Ziteng Cui1,2
, Liuzhuozheng Li1
Gengjia Chang2
, Jun Liu2
, Ziying Gu1
, Dong Li2
, Xuangeng Chu1,2
Lin Gu3
, Tatsuya Harada1,4
1The University of Tokyo 2
I2WM 3Tohoku University 4RIKEN AIP
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. FluxFlow: Conservative Flow-Matching for
Astronomical Image Super-Resolution
Shuhong Liu1,2, $
, Xining Ge2,
, Ziteng Cui1,2
, Liuzhuozheng Li1
Gengjia Chang2
, Jun Liu2
, Ziying Gu1
, Dong Li2
, Xuangeng Chu1,2
Lin Gu3
, Tatsuya Harada1,4
1The University of Tokyo 2
I2WM 3Tohoku University 4RIKEN AIP
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. FluxFlow: Conservative Flow-Matching for
Astronomical Image Super-Resolution
Shuhong Liu1,2, $
, Xining Ge2,
, Ziteng Cui1,2
, Liuzhuozheng Li1
Gengjia Chang2
, Jun Liu2
, Ziying Gu1
, Dong Li2
, Xuangeng Chu1,2
Lin Gu3
, Tatsuya Harada1,4
1The University of Tokyo 2
I2WM 3Tohoku University 4RIKEN AIP
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.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 1 ---
FluxFlow: Conservative Flow-Matching for Astronomical Image Super-Resolution Shuhong Liu1,2, Xining Ge2,, Ziteng Cui1,2, Liuzhuozheng Li1 Gengjia Chang2, Jun Liu2, Ziying Gu1, Dong Li2, Xuangeng Chu1,2 Lin Gu3, Tatsuya Harada1,4 1The University of Tokyo 2 I2WM 3Tohoku University 4RIKEN AIP 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 <ref:2605.03749#pg2> 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 <ref:2605. p. 0 FluxFlow: Conservative Flow-Matching for
Astronomical Image Super-Resolution
Shuhong Liu1,2, $
, Xining Ge2,
, Ziteng Cui1,2
, Liuzhuozheng Li1
Gengjia Chang2
, Jun Liu2
, Ziying Gu1
, Dong Li2
, Xuangeng Chu1,2
Lin Gu3
, Tatsuya Harada1,4
1The University of Tokyo 2
I2WM 3Tohoku University 4RIKEN AIP
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. FluxFlow: Conservative Flow-Matching for
Astronomical Image Super-Resolution
Shuhong Liu1,2, $
, Xining Ge2,
, Ziteng Cui1,2
, Liuzhuozheng Li1
Gengjia Chang2
, Jun Liu2
, Ziying Gu1
, Dong Li2
, Xuangeng Chu1,2
Lin Gu3
, Tatsuya Harada1,4
1The University of Tokyo 2
I2WM 3Tohoku University 4RIKEN AIP
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. 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 Wienerregularized test-time correction to suppress hallucination sources while preserving
recovered detail.Where the paper says it
Where the paper says it
Where the paper says it
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 <ref:2605.03749#pg2> 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 <ref:2605.03749#pg2>
During training, it uses spatial reliability and source-importance maps to focus learning on scientifically focused regions while suppressing unstable gradients from artifacts <ref:2605.03749#pg2> The overall objective is defined as LWFM = Et, x0, x1 [P p Wf ⊙ S] vθ(xt, t) − ut2 p P p Wf ⊙ S <ref:2605. p. 4 Overall Objective. Multiplying the two maps yields a single spatially varying weight that encodes
both reliability and importance, and is applied as a per-image normalized objective:
LWFM = Et, x0, x1
P
p
Wf ⊙ S
p
vθ(xt, t, c) − ut
2
p P
p
Wf ⊙ S
p
, (5)
where c denotes the conditioning upsampled DESI observation specified in Appendix D, ⊙ is
element-wise multiplication, p indexes spatial pixel locations, and the denominator P
p
(Wf ⊙ S)p is
recomputed per training pair so that each pair contributes a gradient signal of comparable magnitude
irrespective of the absolute scale of Wf ⊙ S.Where the paper says it
During inference, it enforces measurement consistency with the observed low-resolution image by back-projecting residuals through a Wiener-regularized inverse of the forward operator to suppress hallucination sources while preserving recovered detail <ref:2605.03749#pg2> This correction is implemented using Equation (11) which replaces the PSF adjoint in A⊤ with the Wiener kernel W(k) <ref:2605. p. 6 With these strict adjoints,
Equation (9) becomes the exact gradient step on
1
2
∥A(x) − y∥
2
, and Equation (11) replaces H⊤
PSF
in this step with the Wiener kernel W. This prevents the cumulative PSF widening of repeated adjoint projections while retaining the hallucinationsuppression property.Where the paper says it
Experiments on our curated dataset demonstrate that FluxFlow achieves superior photometric fidelity and scientific accuracy over both recent regression-based and generative approaches <ref:2605.03749#pg2> Table 2 shows quantitative results at ×2 and ×4 on our DESI–HST super-resolution dataset, where FluxFlow obtains PSNR competitive with regression baselines, within 0. p. 7 1 Quantitative and Qualitative Evaluation
Table 2 shows quantitative results at ×2 and ×4 on our DESI–HST dataset. FluxFlow obtains
PSNR competitive with regression baselines, within 0. 1 Quantitative and Qualitative Evaluation
Table 2 shows quantitative results at ×2 and ×4 on our DESI–HST dataset. FluxFlow obtains
PSNR competitive with regression baselines, within 0. FluxFlow obtains
PSNR competitive with regression baselines, within 0. 1 dB of the highest value at both scales, while
achieving the best SSIM and Flux-L1 among all compared methods. FluxFlow obtains
PSNR competitive with regression baselines, within 0. 1 dB of the highest value at both scales, while
achieving the best SSIM and Flux-L1 among all compared methods. FluxFlow obtains
PSNR competitive with regression baselines, within 0. 1 dB of the highest value at both scales, while
achieving the best SSIM and Flux-L1 among all compared methods. Figure 7 presents qualitative comparisons at ×2. Regression-based methods
tend to oversmooth high-frequency structures, whereas FluxFlow recovers fine details including
the diffraction spikes around bright stars and the sharp cores of point sources.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Hallucinations are suppressed rather than eliminated, and scientific accuracy still falls short of what precision astronomical measurements demand <ref:2605.03749#pg2> Future work will extend FluxFlow to multi-band joint reconstruction and uncertainty-calibrated sampling to improve the SNR <ref:2605.03749#pg2> The paper is supported by a comprehensive set of references covering flow matching, diffusion models, and astronomical image restoration techniques <ref:2605.03749#pg2> The final comparison across all four test cases shows FluxFlow recovers diffraction patterns, faint background structures, and source morphology that regression baselines oversmooth and that generative baselines either speckle or oversharpen <ref:2605.03749#pg2> The paper is supported by a comprehensive set of references covering flow matching, diffusion models, and astronomical image restoration techniques <ref:2605.03749#pg2> The final comparison across all four test cases shows FluxFlow recovers diffraction patterns, faint background structures, and source morphology that regression baselines oversmooth and that generative baselines either speckle or oversharpen <ref:2605.03749#pg2>
--- Page 2 ---
Ideal Source Diffraction-limited Airy Disk Seeing-limited Observation Figure 2 Illustration of point source degradation from ideal (left) to diffraction-limited (center) to seeing-limited (right), with 2D intensity images and 1D cross-sectional profiles <ref:2605.03749#pg2> This distinction is especially important in ground-to-space image super-resolution <ref:2605. p. 1 Ideal Source Diffraction-limited Airy Disk Seeing-limited Observation
Figure 2: Illustration of point source degradation from ideal (left) to diffraction-limited (center) to
seeing-limited (right), with 2D intensity images and 1D cross-sectional profiles. This distinction is especially important in ground-to-space image super-resolution. Ideal Source Diffraction-limited Airy Disk Seeing-limited Observation
Figure 2: Illustration of point source degradation from ideal (left) to diffraction-limited (center) to
seeing-limited (right), with 2D intensity images and 1D cross-sectional profiles. This distinction is especially important in ground-to-space image super-resolution. This distinction is especially important in ground-to-space image super-resolution. As illustrated
in Figure 2, Ground-based observations are typically seeing-limited, where atmospheric turbulence
broadens the PSF and imposes an angular-resolution floor that often dominates over the detector
sampling scale. This distinction is especially important in ground-to-space image super-resolution. As illustrated
in Figure 2, Ground-based observations are typically seeing-limited, where atmospheric turbulence
broadens the PSF and imposes an angular-resolution floor that often dominates over the detector
sampling scale. As illustrated
in Figure 2, Ground-based observations are typically seeing-limited, where atmospheric turbulence
broadens the PSF and imposes an angular-resolution floor that often dominates over the detector
sampling scale. In contrast, space-based telescopes such as the Hubble Space Telescope (HST)
are free from atmospheric seeing and can approach the diffraction limit of their optics, yielding
substantially sharper images. As illustrated
in Figure 2, Ground-based observations are typically seeing-limited, where atmospheric turbulence
broadens the PSF and imposes an angular-resolution floor that often dominates over the detector
sampling scale. In contrast, space-based telescopes such as the Hubble Space Telescope (HST)
are free from atmospheric seeing and can approach the diffraction limit of their optics, yielding
substantially sharper images. However, its small field of view and heavily oversubscribed observing
time restrict such image quality to a small fraction of the sky, creating a persistent gap between
wide-field ground-based surveys and high-resolution space-based observations. Bridging this gap would enable high-resolution analysis over survey scales that are inaccessible
to space telescopes alone. However, its small field of view and heavily oversubscribed observing
time restrict such image quality to a small fraction of the sky, creating a persistent gap between
wide-field ground-based surveys and high-resolution space-based observations. Bridging this gap would enable high-resolution analysis over survey scales that are inaccessible
to space telescopes alone.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Ideal Source Diffraction-limited Airy Disk Seeing-limited Observation Figure 2 Illustration of point source degradation from ideal (left) to diffraction-limited (center) to seeing-limited (right), with 2D intensity images and 1D cross-sectional profiles <ref:2605.03749#pg2> This distinction is especially important in ground-to-space image super-resolution <ref:2605.03749#pg2> As illustrated in Figure 2, Ground-based observations are typically seeing-limited, where atmospheric turbulence broadens the PSF and imposes an angular-resolution floor that often dominates over the detector sampling scale <ref:2605.03749#pg2> In contrast, space-based telescopes such as the Hubble Space Telescope (HST) are free from atmospheric seeing and can approach the diffraction limit of their optics, yielding substantially sharper images <ref:2605.03749#pg2> Bridging this gap would enable high-resolution analysis over survey scales that are inaccessible to space telescopes alone <ref:2605.03749#pg2>
The authors construct the DESI–HST Dataset, a real paired ground-to-space super-resolution benchmark comprising 19,500 real co-registered ground-to-space image pairs with real atmospheric PSF variation and observation uncertainty maps <ref:2605.03749#pg2>
The training objective uses spatial reliability and source-importance maps to focus learning on scientifically focused regions while suppressing unstable gradients from artifacts <ref:2605.03749#pg2> The inference correction enforces measurement consistency with the observed low-resolution image by back-projecting residuals through a Wiener-regularized inverse of the forward operator <ref:2605. p. 1 During training, it uses spatial reliability and source-importance maps to focus learning on
scientifically focused regions while suppressing unstable gradients from artifacts. During inference,
it enforces measurement consistency with the observed low-resolution image by back-projecting
residuals through a Wiener-regularized inverse of the forward operator.Where the paper says it
Experiments on our curated dataset demonstrate that FluxFlow achieves superior photometric fidelity and scientific accuracy over both recent regression-based and generative approaches <ref:2605.03749#pg2> Table 2 shows quantitative results at ×2 and ×4 on our DESI–HST super-resolution dataset, where FluxFlow obtains PSNR competitive with regression baselines, within 0.1 dB of the highest value at both scales, while achieving the best SSIM and Flux-L1 among all compared methods <ref:2605.03749#pg2>
The scarcity of real observational data for ground-to-space SR is overcome by constructing the DESI–HST Dataset, the large-scale astronomical SR dataset comprising real-world co-registered ground-to-space image pairs with authentic atmospheric PSF variation and observation uncertainty maps <ref:2605.03749#pg2>
--- Page 3 ---
滚滚长江东逝水 Figure 3 Paired examples from our DESI–HST dataset. Top: DESI cutouts at 128×128 pixels; Bottom: HST observations at 512×512 pixels. p. 2 滚滚长江东逝水
Figure 3: Paired examples from our DESI–HST dataset. Top: DESI cutouts at 128×128 pixels;
Bottom: HST observations at 512×512 pixels. 滚滚长江东逝水
Figure 3: Paired examples from our DESI–HST dataset. Top: DESI cutouts at 128×128 pixels;
Bottom: HST observations at 512×512 pixels. Top: DESI cutouts at 128×128 pixels;
Bottom: HST observations at 512×512 pixels. All pairs are pixel-aligned. Top: DESI cutouts at 128×128 pixels;
Bottom: HST observations at 512×512 pixels. All pairs are pixel-aligned.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Figure 3 Paired examples from our DESI–HST dataset. Top: DESI cutouts at 128×128 pixels; Bottom: HST observations at 512×512 pixels. All pairs are pixel-aligned <ref:2605.03749#pg2>
The training objective uses spatial reliability and source-importance maps to focus learning on scientifically focused regions while suppressing unstable gradients from artifacts <ref:2605.03749#pg2> The inference correction enforces measurement consistency with the observed low-resolution image by back-projecting residuals through a Wiener-regularized inverse of the forward operator to suppress hallucination sources while preserving recovered detail <ref:2605. p. 1 During inference,
it enforces measurement consistency with the observed low-resolution image by back-projecting
residuals through a Wiener-regularized inverse of the forward operator. Together, these designs
suppress the generation of hallucination sources while preserving the fine structure.Where the paper says it
--- Page 4 ---
Table 1 Comparison of astronomical SR datasets G and S denote ground- and space-based imaging Dataset Size Real Flux SR-Type Miao et al. [2024] 2,000 × G→G Shan et al. p. 3 Dataset Size Real Flux SR-Type
Miao et al. [2024] 2,000 ✓ × G→G
Shan et al. [2024] 2,000 ✓ × G→G
Shan et al. [2025] 9,383 ✓ × G→G
Song et al. [2025] 9,383 ✓ × G→G
Song et al. [2024] 1,597 ✓ × S→G
Reddy et al. [2024] 1,597 ✓ × S→G
Reddy et al. [2024]
1
2,880 ✓ × G→S
Luo et al. Dataset Size Real Flux SR-Type
Miao et al. [2024] 2,000 ✓ × G→G
Shan et al. [2025] 17,000 ✓ × G→G
Li et al. [2022] 14,604 ✓ × G→G
Wu et al. [2022] 14,604 ✓ × G→G
Wu et al. [2025] 54,738 ✓ ✓ S→S
Zhang et al. 1 Quantitative and Qualitative Evaluation
Table 2 shows quantitative results at ×2 and ×4 on our DESI–HST dataset. FluxFlow obtains
PSNR competitive with regression baselines, within 0.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Table 1 Comparison of astronomical SR datasets G and S denote ground- and space-based imaging Dataset Size Miao et al. [2605.037491] 2,000 × G→G Shan et al. p. 3 Dataset Size Real Flux SR-Type
Miao et al. [2024] 2,000 ✓ × G→G
Shan et al. [2024] 2,000 ✓ × G→G
Shan et al. [2025] 9,383 ✓ × G→G
Song et al.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Conservative training objective: The training objective is modified to include spatial reliability and source-importance maps to
focus learning on scientifically focused regions while suppressing unstable gradients from artifacts.
This ensures that the model learns from reliable data while ignoring noise or artifacts that lead to unstable gradients. -
Training-free test-time correction: A training-free Wiener regularized test-time correction is incorporated to
suppress hallucination sources while preserving recovered detail.
This mechanism enforces measurement consistency by back-projecting residuals through a Wiener regularized inverse of the forward operator, which prevents the generation of artifacts in inference. -
Physically informed weighting: The objective function incorporates spatially varying weights that encode both data reliability and scientific importance, yielding a single spatially varying weight that
encodes both reliability and importance.
This ensures that the loss is weighted per-pixel based on inverse variance (Wf) and source mask (S). -
Measurement-consistent flow sampling: The inference procedure utilizes a measurement-consistent flow sampling scheme where the residual is projected back to high resolution using a Wiener regularized approximate adjoint of the rank-deficient operator A. This enforces consistency with the observation by penalizing trajectories that exit the set C(y), which is
the set of high-resolution images x consistent with a given DESI observation y.
-
High fidelity photometric and scientific accuracy: The system achieves superior photometric fidelity and scientific accuracy over baselines, as demonstrated by achieving competitive PSNR with regression methods while yielding the
best SSIM and Flux-L1 among all compared methods.
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. Project webpage: https://i2wm.github.io/fluxflow/.
Sources
- Beyond Model Design: Data-Centric Training and Self-Ensemble for Gaussian Color Image Denoising
- Training-Free Model Ensemble for Single-Image Super-Resolution via Strong-Branch Compensation
- Diffusion Posterior Sampling for General Noisy Inverse Problems
- Training Flow Matching: The Role of Weighting and Parameterization
- Dual-Branch Remote Sensing Infrared Image Super-Resolution
- CLIP-Guided Data Augmentation for Night-Time Image Dehazing
- Classifier-Free Diffusion Guidance
- One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation
- Exploring Time Conditioning in Diffusion Generative Models from Disjoint Noisy Data Manifolds
- Back to Basics: Let Denoising Generative Models Denoise
- Flow Matching for Generative Modeling
- I$^2$SB: Image-to-Image Schr\"odinger Bridge
- RealX3D: A Physically-Degraded 3D Benchmark for Multi-view Visual Restoration and Reconstruction
- I2-NeRF: Learning Neural Radiance Fields Under Physically-Grounded Media Interactions
- NTIRE 2026 3D Restoration and Reconstruction in Real-world Adverse Conditions: RealX3D Challenge Results
- Denoising the Deep Sky: Physics-Based CCD Noise Formation for Astronomical Imaging
- Fine-grained Image Editing by Pixel-wise Guidance Using Diffusion Models
- DIAMOND: Directed Inference for Artifact Mitigation in Flow Matching Models
- Flower: A Flow-Matching Solver for Inverse Problems
- Generative AI for image reconstruction in Intensity Interferometry: a first attempt
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs