AstroCompress: A benchmark dataset for multi-purpose compression of astronomical data

Joshua S. Bloom; Peter Xiangyuan Ma; Rithwik Sudharsan; Ruihan Yang; Stephan Mandt; Tuan Truong; Yibo Yang

REVIEW 3 major objections 5 minor 51 references

AstroCompress: A benchmark dataset for multi-purpose compression of astronomical data

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AstroCompress benchmarks seven lossless compressors on 320 GB of 16-bit astronomical imagery and finds that minimally adapted neural codecs match or beat the best classical codecs on single-frame data.

desk verdict The corpus is the contribution; the neural-vs-classical ranking is plausible but still an estimate until someone runs an actual arithmetic coder. read the letter →

arxiv 2506.08306 v1 pith:53OII3RG submitted 2025-06-10 cs.AI astro-ph.IM

Tuan Truong , Rithwik Sudharsan , Yibo Yang , Peter Xiangyuan Ma , Ruihan Yang , Stephan Mandt , Joshua S. Bloom This is my paper

classification cs.AIastro-ph.IM

keywords astronomicaldatacompressionlosslessneuralbenchmarkdataset16-bitimagingobservatorydownlinklikelihood-basedcodecsJPEG-XL

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

AstroCompress is a public benchmark of roughly 320 GB of 16-bit astronomical imaging, organized into five datasets that span ground-based and space-based observations, multi-wavelength filters, and time-series cubes. The paper runs seven lossless compressors—three neural and four classical—over a fixed train/test split and reports compression ratios. Its central claim is that neural compression, with minimal changes for 16-bit pixels, can match or surpass the best classical codecs on single-frame telescope images, with concrete wins on ground-based Keck data and JWST first frames. A diffusion-based likelihood model suggests the neural results still have considerable headroom, while classical codecs remain ahead when frames are stacked in time or wavelength. If correct, this gives observatories with tight downlink budgets a practical reason to invest in learned codecs and recover more science per byte transmitted.

What carries the argument

The load-bearing object is the corpus itself: five collections of 16-bit unsigned integer arrays in 2D, 3D time-series, and 4D time-by-wavelength formats, assembled with non-overlapping sky footprints between training and test splits. The identity that carries the argument is the Shannon-style equivalence between entropy coding and likelihood: compression ratio is reported as uncompressed bit depth divided by the model's negative log-likelihood in bits per pixel, which the paper aligns with arithmetic coding performance. This lets all seven methods be scored on one scale and lets a generative model's likelihood stand in for an actual compressed file. Around that identity, the paper defines the adaptation trick for 16-bit data (treat each pixel as two 8-bit channels, most- and least-significant bytes, or widen the likelihood's bin count) and the fixed patch-based evaluation protocol.

What would settle it

Run a real arithmetic coder over the trained neural models on the full test images and compare on-disk byte counts with the negative-log-likelihood ratios; if the neural ratios on Keck or JWST-2D fall below JPEG-XL max once entropy coding overhead is included, the 'match or surpass' claim fails at the file level.

Watch

Extended reading notes

Core claim

The paper's central claim is that lossless neural compression, minimally adapted from natural-image models to 16-bit telescope data, can match or surpass the best classical codecs on real astronomical imaging. In the single-frame benchmark, Integer Discrete Flows and PixelCNN++ beat JPEG-XL in maximum-effort mode on ground-based Keck images (2.04 and 2.08 versus 2.01) and on JWST first-frame images (1.44 versus 1.38); VDM, a diffusion-based likelihood model, exceeds all classical codecs on Hubble images (3.33 versus 3.26), indicating headroom. The neural superiority is not universal: JPEG-XL max still leads on the LCO and SDSS single-frame sets, and on temporally or spectrally stacked data every classical codec except the weakest gains a larger boost than the neural ones. The paper interprets this as evidence that neural codecs are viable for telescope downlinks today, with the main wins in single-frame space-based and ground-based imaging.

Load-bearing premise

The neural compression ratios are computed from model negative log-likelihoods rather than actual entropy-coded files, so the conclusion assumes that arithmetic coding overhead and any gap between likelihood and on-disk bit counts are negligible.

Editorial extensions

If this is right

A fractional improvement in lossless ratio converts directly into additional transmitted science; the paper quotes tens of millions of dollars per one percent gain for a single flagship space telescope and billions across the observatory fleet.
The VDM likelihood results imply that current neural benchmarks sit well below the achievable ceiling, so more practical diffusion or flow codecs should push ratios higher without changing the corpus.
Single-frame neural wins do not yet extend to temporally or spectrally stacked data, where classical codecs still lead; cross-frame and cross-filter modeling is the immediate open problem.
Because background noise dominates bit allocation, codecs that model noise accurately will gain the most, and the paper's noise analysis gives a concrete target for lossy or near-lossless designs.
The strong generalization from diverse training sets suggests one multi-purpose compressor trained on mixed data could serve many instruments, potentially succeeding JPEG-XL as a general astronomy codec.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: if the likelihood-to-disk gap closes with efficient arithmetic coding, the practical deployment case strengthens immediately, because the paper's runtime table already shows IDF is about ten times faster than JPEG-XL max while beating it on two datasets.
Beyond the paper: the bitrate-versus-SNR step function, where background pixels sit near three bits per pixel and sources jump to roughly ten, suggests a near-lossless codec that spends bits only on sources could cut on-disk size far below lossless ratios; the paper stops at lossless.
Beyond the paper: the zero-overlap train/test construction, with hierarchical clustering on angular separation and spherical-polygon checks, is a reusable template for any science benchmark where nearby observations share structure; leakage between adjacent fields would otherwise inflate compression scores.
Beyond the paper: the Keck-trained model beating an SDSS-trained model on SDSS test data indicates dataset diversity, not just size, drives transfer; future instrument teams could collect calibration data across filters and conditions specifically to train a universal codec.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Punchline: the corpus is the contribution; the neural-vs-classical ranking in Table 1 is plausible but still an estimate until someone runs an actual arithmetic coder.

The five datasets (four new, one legacy) are a genuinely useful resource: raw 16-bit unsigned images from Keck, HST, JWST, and SDSS Stripe 82, including 3D and 4D cubes, with careful sky-overlap filtering so train and test don't share sky regions. The HuggingFace release comes with code, download scripts, and a reproducible split. That alone makes the paper worth engaging with, and it fills a real gap—previous compression corpora were 8-bit, ground-based-only, or much smaller.

The benchmark is a solid first cut, not a definitive ranking. PixelCNN++ and IDF beat JPEG-XL on some datasets (Hubble, Keck, JWST-2D) and lose on others (LCO, SDSS-2D). The authors are appropriately careful to call this 'match or surpass' rather than 'dominates,' and the stress-test is right that the central existence claim survives the main caveat. The caveat is real: neural numbers are negative log-likelihoods, not measured file sizes (Sec. 4.2), and the paper admits in Sec. 6.1 that no entropy coding was run. For autoregressive and flow models the overhead is usually well under 0.1 bits per pixel, while the winning margins are 0.12–0.20, so the direction is believable. Still, for a benchmark meant to guide real deployments, I'd want actual bitstreams on at least one dataset, error bars in Table 1, and a check that the best-of-two-variant reporting in Sec. C.1 doesn't inflate the neural numbers. Those are precision problems, not fatal ones.

The qualitative analyses—compressibility vs. background noise, SNR vs. bitrate—are nice and consistent with older work like Pence et al. (2009). The runtime table is transparent that neural times are likelihood evaluation, not full encoding.

Weakest section is the introduction's economic extrapolation. The '$43M for a 1% improvement on Roman' is simple multiplication, presented as motivation. It's clearly labeled as such, but it would be easy for readers to over-weight it.

Bottom line: this is a dataset and benchmark paper, and the dataset is worth having. The numbers are preliminary but the right next step is clear: run the coder, report error bars, and then the comparison becomes solid. I'd send this to peer review, and I'd cite the dataset in work on high bit-depth compression. I'd also want the authors to close the measurement gap in a revision.

Referee Report

3 major / 5 minor

Summary. The paper introduces AstroCompress, a ~320 GB corpus of uint16 astronomical imaging data drawn from five sources (Keck, Hubble, JWST, SDSS, and a legacy LCO subset), with careful train/test separation based on sky-overlap filtering. It benchmarks three practical neural lossless compressors (IDF, L3C, PixelCNN++) plus VDM as a likelihood-based upper bound against four classical codecs (JPEG-XL in two modes, JPEG-LS, JPEG-2000, RICE). The main empirical claim is that neural methods can match or even surpass the best classical codec on some datasets, specifically Keck and JWST-2D, and the paper also analyzes noise/bitrate correlations and cross-dataset generalization.

Significance. If the quantitative comparison is validated, AstroCompress is a valuable public benchmark and the results provide an existence proof that learned lossless codecs are competitive for telescope data, motivating further work toward deployable neural compression in astronomy. The corpus is large, spans diverse observing modes, and the explicit non-overlapping-sky filtering is a careful design choice that supports honest generalization measurements. The release of data-loading code, baseline evaluation scripts, and per-image CSV results is a genuine service to the community. The central claim, however, rests on likelihood-based bitrate estimates rather than measured bitstreams, and the small margins on key datasets make this assumption material.

major comments (3)

[§4.2, Table 1; §6.1] The neural compression ratios in Table 1 are computed as 16 divided by the negative log-likelihood assigned by the model, not from actual encoded bitstreams, and §6.1 confirms that no entropy coding was run. For the two datasets where practical neural methods beat JPEG-XL (max), the margins are small (Keck: IDF 2.04 vs 2.01; JWST-2D: IDF and PixelCNN++ 1.44 vs 1.38). Arithmetic coding overhead, while expected to be small, is not quantified, so the 'surpass' wording is not yet established to the precision claimed. Please report actual arithmetic-coded file sizes for at least these close comparisons, or provide a per-dataset measurement of the overhead between the likelihood estimate and the real bitstream.
[Table 1, §4.2] No measure of variability is provided for any compression ratio in Table 1. Figure 2 shows substantial image-to-image scatter in compression ratio as a function of exposure time, so the reported averages may hide overlaps between methods. Please provide standard deviations or confidence intervals across test images (and, where applicable, across training runs), and state whether the Keck and JWST-2D differences between neural methods and JPEG-XL (max) are statistically significant.
[§4.2, §C.1, Table 5] The paper reports the better of two model variants (direct 16-bit input vs. 8-bit MSB/LSB channel-splitting) after evaluating both on the test set. Table 5 shows that the two variants differ dramatically on some datasets (PixelCNN++ LCO 1.41 vs 2.02; SDSS-2D 3.35 vs 1.84). This test-set model selection inflates reported ratios and makes the comparison to classical codecs, which are not similarly cherry-picked, unfair. The variant should be chosen on the validation split and fixed before test evaluation.

minor comments (5)

[Abstract] The phrase 'any improvements in lossless data compression has the potential scale' appears to be missing 'to'; it should read 'has the potential to scale'.
[§4.3] The sentence 'The estimated compression performance of VDM significantly surpass all existing methods' has a subject-verb agreement error; 'surpass' should be 'surpasses'.
[Table 4] In the LCO row, the image size '136×2112' looks like a typo; please check the actual dimensions, as the stated total pixel count and dataset size imply a different value.
[§6.1] The statement 'Our relatively inefficient entropy coding implementation shows a negligible overhead' is hard to reconcile with the earlier statement that no entropy coding was run; please clarify which entropy coder was used and for which results the overhead was measured.
[§4.1] The assertion that 'the latest algorithm in current use is Rice' is stronger than the cited literature supports, since CCSDS has published newer lossless multispectral/hyperspectral recommendations; a softer phrasing or a more recent citation would be preferable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark compares trained neural models and published classical codecs on held-out test data, and the likelihood-to-bitrate mapping is a standard information-theoretic proxy rather than a forced identity.

full rationale

The central claim that neural lossless compression can match or surpass classical codecs is an empirical existence claim supported by Table 1, where neural and non-neural methods are evaluated on the same held-out images. The compression ratios for neural models are computed as uncompressed bit depth divided by the negative log-likelihood assigned by the model (Sec. 4.2); this is the Shannon/arithmetic-coding equivalence, not a self-definitional reduction, and the paper explicitly acknowledges that no full entropy-coding pass was run (Sec. 6.1), which is a measurement limitation rather than a circular step. The classical baselines are independent, publicly available codecs (libjxl, charls, OpenJPEG, cfitsio) run with default settings, and the authors validate their classical results against published LCO numbers from Maireles-Gonzalez et al. (2023). Train/test separation with conservative sky-overlap filtering (Sec. A.4) prevents the neural results from being fits to the test inputs. The few self-citations (e.g., Yang et al. 2023b, Yang & Mandt 2023) appear only in background motivation and do not carry the load of any benchmark conclusion; no uniqueness theorem or ansatz is imported from the authors' prior work to force the chosen methods. Reporting the better of two input-format variants (Sec. C.1) is a model-selection decision that could affect individual rows, but it does not make the results reduce by construction to the paper's own inputs. Overall, no circular step was identified.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No free parameters were fitted to produce the benchmark results; the neural models have learned weights, but these are standard training artifacts, not hand-tuned numbers used to force the conclusion. The key axioms are the likelihood-to-bitrate equivalence, Shannon's theorem, the sky-overlap split, and the representativeness of the selected codecs. No new physical or conceptual entities are introduced.

assumptions (4)

domain assumption The negative log-likelihood assigned by a neural model equals the number of bits an actual entropy coder would produce on disk.
Used to report neural compression ratios in Table 1 without running arithmetic coding (Sec. 4.2 and 6.1). The paper calls the overhead negligible but does not provide measured bitstreams.
standard math Shannon's source coding theorem connects likelihood estimation to lossless compression rate.
Invoked in Sec. 6 to justify using likelihood scores as compression costs.
domain assumption The sky-footprint non-overlap procedure guarantees the test set is not implicitly seen in training.
Used for the train/test split (Sec. 3 and Appendix A.4). Keck had no WCS-based check and Legacy no filtering, so the guarantee is weaker for those datasets.
domain assumption The selected neural and classical algorithms are representative of the current practical state of the art.
Sec. 4.1 claims the comparison includes all practical state-of-the-art algorithms; this is a domain judgment and could be challenged as new codecs appear.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AstroCompress: A benchmark dataset for multi-purpose compression of astronomical data." pith.science (2026). https://pith.science/paper/53OII3RG

@misc{pith2026250608306,
  author       = {Pith},
  title        = {Pith review of: AstroCompress: A benchmark dataset for multi-purpose compression of astronomical data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/53OII3RG}},
  note         = {Machine review of arXiv:2506.08306}
}

read the original abstract

The site conditions that make astronomical observatories in space and on the ground so desirable -- cold and dark -- demand a physical remoteness that leads to limited data transmission capabilities. Such transmission limitations directly bottleneck the amount of data acquired and in an era of costly modern observatories, any improvements in lossless data compression has the potential scale to billions of dollars worth of additional science that can be accomplished on the same instrument. Traditional lossless methods for compressing astrophysical data are manually designed. Neural data compression, on the other hand, holds the promise of learning compression algorithms end-to-end from data and outperforming classical techniques by leveraging the unique spatial, temporal, and wavelength structures of astronomical images. This paper introduces AstroCompress: a neural compression challenge for astrophysics data, featuring four new datasets (and one legacy dataset) with 16-bit unsigned integer imaging data in various modes: space-based, ground-based, multi-wavelength, and time-series imaging. We provide code to easily access the data and benchmark seven lossless compression methods (three neural and four non-neural, including all practical state-of-the-art algorithms). Our results on lossless compression indicate that lossless neural compression techniques can enhance data collection at observatories, and provide guidance on the adoption of neural compression in scientific applications. Though the scope of this paper is restricted to lossless compression, we also comment on the potential exploration of lossy compression methods in future studies.

Figures

Figures reproduced from arXiv: 2506.08306 by the authors.

**Figure 3.** From left to right, an example SDSS-2D image: raw image, SNR heatmap, and PixelCNN++ [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Left: SNR heatmap of an example image ( [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Correlation between background noise variation and compressibility. Each point represents [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 29 canonical work pages

[1]

cosmic-ray [CR] hits

A.3.3 SBI-16-2D (HUBBLE) This dataset is based on data from the Wide Field Channel (WFC) instrument of the Advanced Camera for Surveys (ACS) onboard the Hubble Space Telescope (HST). Unlike the GBI-16-2D dataset, all observations in SBI-16-2D are obtained in space and with the same bandpass filter (F606W), providing a more uniform point-spread function ac...

work page
[2]

Borucki, David Koch, Gibor Basri, Natalie Batalha, Timothy Brown, Douglas Caldwell, John Caldwell, Jørgen Christensen-Dalsgaard, William D

William J. Borucki, David Koch, Gibor Basri, Natalie Batalha, Timothy Brown, Douglas Caldwell, John Caldwell, Jørgen Christensen-Dalsgaard, William D. Cochran, Edna DeV ore, Edward W. Dunham, Andrea K. Dupree, Thomas N. Gautier, John C. Geary, Ronald Gilliland, Alan Gould, Steve B. Howell, Jon M. Jenkins, Yoji Kondo, David W. Latham, Geoffrey W. Marcy, Sø...

work page 2025
[3]

reversible color transforms

and then randomly sampled fields and determined the center of the field sky positions. For each such field, we queried the SDSS Stripe 82 database for other images (regardless of quality) that overlap the field center and downloaded those images. Since field images from different runs are not aligned in RA, we also downloaded the two fields immediately ad...

work page 2023
[5]

R” and “G

• To implement approach (1), we make the following modifications to support 16-bit input directly: ForIDF, we change the input scaling coefficient from28 to 216, so that it models the set Z/65536 (instead of Z/256). ForL3CandPixelCNN++, which both use a discretized logistic mixture likelihood model (Salimans et al., 2017), we increase the number of bins f...

work page 2019
[6]

exascale

24 Published as a conference paper at ICLR 2025 Model / Dataset L3C-OpenImages (zero-padded) L3C-ImageNet64 (zero-padded) L3C-OpenImages (duplicated) L3C-ImageNet64 (duplicated) IDF-Cifar10 (zero-padded) IDF-Cifar10 (duplicated) LCO 1.78 2.16 2.04 2.05 1.72 1.84 Keck 1.26 1.44 1.27 1.38 0.66 0.59 Hubble 1.77 2.18 1.88 1.90 0.59 1.08 JWST 1.08 1.19 1.06 1....

work page 2025
[7]

Neural data compression for physics plasma simulation

Jong Choi, Michael Churchill, Qian Gong, Seung-Hoe Ku, Jaemoon Lee, Anand Rangarajan, Sanjay Ranka, Dave Pugmire, CS Chang, and Scott Klasky. Neural data compression for physics plasma simulation. InNeural Compression: From Information Theory to Applications–Workshop@ ICLR 2021,

work page 2021
[10]

Accessed: 2024-11-28

URL https://sites.research.google/ neural-mapping/. Accessed: 2024-11-28. Hannah Gulick, Jessica Lu, Steve Beckwith, Joshua Bloom, and Guy Nir. CubeSat for Rapid IR and Optical Surveys. In44th COSPAR Scientific Assembly. Held 16-24 July, volume 44, pp. 1985, July

work page 2024
[12]

Jenkins and Jeb Dunnuck

12 Published as a conference paper at ICLR 2025 Jon M. Jenkins and Jeb Dunnuck. The little photometer that could: technical challenges and science results from the Kepler Mission. In Howard A. MacEwen and James B. Breckinridge (eds.),UV/Optical/IR Space Telescopes and Instruments: Innovative Technologies and Concepts V, volume 8146, pp. 814602. Internatio...

work page 2025

Show all 51 references

[14]

Accessed: 2024-11-28

URL https://jwst-docs.stsci.edu/. Accessed: 2024-11-28. S. M. Kahn, N. Kurita, K. Gilmore, M. Nordby, P. O’Connor, R. Schindler, J. Oliver, R. Van Berg, S. Olivier, V . Riot, P. Antilogus, T. Schalk, M. Huffer, G. Bowden, J. Singal, and M. Foss. Design and development of the 3...

2024
[15]

Nodirkhuja Khujaev, Roman Tsoy, and Seungryul Baek

doi: 10.1117/12.857920. Nodirkhuja Khujaev, Roman Tsoy, and Seungryul Baek. AstroYOLO: Learning Astronomy Multi- tasks in a Single Unified Real-Time Framework. InMachine Learning and the Physical Sciences Workshop, NeurIPS,

doi
[16]

Anji Liu, Stephan Mandt, and Guy Van den Broeck

doi: 10.1093/mnras/sty3217. Anji Liu, Stephan Mandt, and Guy Van den Broeck. Lossless compression with probabilistic circuits. InProceedings of the International Conference on Learning Representations (ICLR),

doi
[17]

URL https://mast.stsci. edu/. Accessed: 2024-11-28. Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Practical full resolution learned lossless image compression. InProceedings of the IEEE Conference on Computer Vision and Pattern Recogniti...

2024
[18]

Vct: A video compression transformer.arXiv preprint arXiv:2206.07307,

Fabian Mentzer, George Toderici, David Minnen, Sung-Jin Hwang, Sergi Caelles, Mario Lucic, and Eirikur Agustsson. Vct: A video compression transformer.arXiv preprint arXiv:2206.07307,

arXiv
[19]

com/watch?v=84XkOlBhOas

URLhttps://www.youtube. com/watch?v=84XkOlBhOas. 13 Published as a conference paper at ICLR 2025 NASA. A Look at the Numbers as NASA’s Hubble Space Telescope Enters its 25th Year,

2025
[21]

nasa.gov/wp-content/uploads/2024/07/ig-24-014.pdf

URL https://oig. nasa.gov/wp-content/uploads/2024/07/ig-24-014.pdf. NIH. Genomic Data Science,

2024
[22]

Accessed: 2024-11-28

URL https://www.genome.gov/about-genomics/ fact-sheets/Genomic-Data-Science. Accessed: 2024-11-28. J. B. Oke, J. G. Cohen, M. Carr, J. Cromer, A. Dingizian, F. H. Harris, S. Labrecque, R. Lucinio, W. Schaal, H. Epps, and J. Miller. The Keck Low-Resolution Imaging Spectrometer....

2024
[25]

photutils

doi: 10.1051/0004-6361/ 201015362. photutils. photutils, An Astropy Package for Photometry,

doi
[26]

readthedocs.io/en/stable/

URL https://photutils. readthedocs.io/en/stable/. Accessed: 2024-11-28. Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. InInternational conference on machine learning, pp. 1530–1538. PMLR,

2024
[28]

Aurko Roy, Mohammad Saffar, Ashish Vaswani, and David Grangier

doi: 10.1117/12.856818. Aurko Roy, Mohammad Saffar, Ashish Vaswani, and David Grangier. Efficient content-based sparse attention with routing transformers.Transactions of the Association for Computational Linguistics, 9:53–68,

doi
[29]

Becker, Peter J

Masao Sako, Bruce Bassett, Andrew C. Becker, Peter J. Brown, Heather Campbell, Rachel Wolf, David Cinabro, Chris B. D’Andrea, Kyle S. Dawson, Fritz DeJongh, Darren L. Depoy, Ben Dilday, Mamoru Doi, Alexei V . Filippenko, John A. Fischer, Ryan J. Foley, Joshua A. Frieman, Lluis...

2025
[30]

Tim Salimans and Jonathan Ho

doi: 10.1088/1538-3873/aab4e0. Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512,

arXiv
[31]

Rob Seaman, William Pence, and Rick White

URL https: //arxiv.org/abs/1701.05517. Rob Seaman, William Pence, and Rick White. fpack: FITS Image Compression Program. Astro- physics Source Code Library, record ascl:1010.002, October

arXiv
[34]

Jeyan Thiyagalingam, Mallikarjun Shankar, Geoffrey Fox, and Tony Hey

https://jwst-docs.stsci.edu/ jwst-general-support/jwst-data-volume-and-data-excess [Accessed: May 30, 2024]. Jeyan Thiyagalingam, Mallikarjun Shankar, Geoffrey Fox, and Tony Hey. Scientific machine learning benchmarks.Nature Reviews Physics, 4(6):413–420, April

2024
[36]

Efficient diffusion models for vision: A survey.arXiv preprint arXiv:2210.09292,

Anwaar Ulhaq and Naveed Akhtar. Efficient diffusion models for vision: A survey.arXiv preprint arXiv:2210.09292,

arXiv
[37]

Kai Wang, Yuanchao Bai, Deming Zhai, Daxin Li, Junjun Jiang, and Xianming Liu

doi: 10.1038/s41586-023-06221-2. Kai Wang, Yuanchao Bai, Deming Zhai, Daxin Li, Junjun Jiang, and Xianming Liu. Learning lossless compression for high bit-depth medical imaging. In2023 IEEE International conference on multimedia and expo (ICME), pp. 2549–2554. IEEE,

doi
[38]

Diffusion Models for Probabilistic Deconvolution of Galaxy Images.arXiv e-prints, art

15 Published as a conference paper at ICLR 2025 Zhiwei Xue, Yuhang Li, Yash Patel, and Jeffrey Regier. Diffusion Models for Probabilistic Deconvolution of Galaxy Images.arXiv e-prints, art. arXiv:2307.11122, July

2025 arXiv
[39]

Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang

doi: 10.48550/arXiv.2307.11122. Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications.ACM Computing Surveys, 56(4):1–39, 2023a. Ruihan Yang, Yibo...

doi arXiv
[40]

Improving inference for neural image compression

Yibo Yang, Robert Bamler, and Stephan Mandt. Improving inference for neural image compression. InNeural Information Processing Systems (NeurIPS), 2020, 2020b. Yibo Yang, Stephan Mandt, and Lucas Theis. An introduction to neural data compression.Founda- tions and Trends® in Com...

2020 doi
[42]

Each camera has its own filter wheel allowing for imaging in a variety of different bandpasses

Since LRIS has a dichroic optical element, which splits the incoming beam at a designated wavelength, the data are obtained either with the blue or 18 Published as a conference paper at ICLR 2025 red side camera. Each camera has its own filter wheel allowing for imaging in a v...

2025
[44]

Stripe 82

It is worth noting that the time gap between subsequent observations varies, depending on readout pattern, but is roughly in the realm of twenty seconds to two minutes. For this reason, we expect that most physical conditions present in the field of view should remain static a...

2000
[45]

camcols”) “field

creates (for each of the 6 camera columns, “camcols”) “field” images spanning a similar declination (ie., north-south) and right ascension (RA; ie., east-west). The images served by the legacy SDSS archive3 are 2048×1489 uint16 pixels, calibrated with a world coordinate system...

2022
[49]

For 3D experiments, we increased the base convolution filter size and adjusted the latent channel size to 96 and 8, respectively

to train the model on 2D image data across all datasets. For 3D experiments, we increased the base convolution filter size and adjusted the latent channel size to 96 and 8, respectively. We trained on 32 × 32 patches, with a learning rate of 1×10 −4 and an exponential decay sc...

2021
[128]

duplicated

We train on patches of size 64 × 64 (using Adam and a learning rate of 2×10 −4) and evaluate on patches of size 256 × 256 to give our model additional context at test time. D ADDITIONALDATAEXPLORATIONS What determines how easily images can be compressed?We answer this question...

2009
[256]

L3CWe adopted the implementation from Mentzer et al

We use a learning rate of1×10 −3 and an exponential decay scheduler with rate0.999. L3CWe adopted the implementation from Mentzer et al. (2019) at https://github.com/ fab-jul/L3C-PyTorch. We largely followed the default model configuration provided by (Mentzer et al.,

2019
[1995]

Petr Pata and Jaromir Schindler

doi: 10.1086/133562. Petr Pata and Jaromir Schindler. Astronomical context coder for image compression.Experimental Astronomy, 39:495–512,

doi
[1998]

James Townsend, Tom Bird, and David Barber

doi: 10.1086/311095. James Townsend, Tom Bird, and David Barber. Practical lossless compression with latent variables using bits back coding.arXiv preprint arXiv:1901.04866,

1901 arXiv
[2000]

telescope

doi: 10.1086/301513. 16 Published as a conference paper at ICLR 2025 Supplementary Material A DATASETSUPPLEMENTALS Telescope Dataset ID Location CentralWavelength (Å)BandpassFilter Resolution(arcsec / pix)Image Sizes(pix×px)Total #ArraysTotal DatasetSize (GB)Approx.Total Pixel...

1995 doi
[2002]

doi: 10.1086/324741. STSCI. JWST Data Volume and Data Excess,

doi
[2005]

Chris Stoughton et al

doi: 10.1086/444553. Chris Stoughton et al. Sloan Digital Sky Survey: Early Data Release.Astronomical Journal, 123(1): 485–548, January

doi
[2009]

doi: 10.1086/599023. W. D. Pence, L. Chiappetti, C. G. Page, R. A. Shaw, and E. Stobie. Definition of the Flexible Image Transport System (FITS), version 3.0.AAP, 524:A42, December

doi
[2010]

Hanqun Cao, Cheng Tan, Zhangyang Gao, Yilun Xu, Guangyong Chen, Pheng-Ann Heng, and Stan Z Li

doi: 10.1126/science.1185402. Hanqun Cao, Cheng Tan, Zhangyang Gao, Yilun Xu, Guangyong Chen, Pheng-Ann Heng, and Stan Z Li. A survey on generative diffusion models.IEEE Transactions on Knowledge and Data Engineering,

doi
[2011]

URLhttps://doi.org/10.1117/12.897767

doi: 10.1117/12.897767. URLhttps://doi.org/10.1117/12.897767. JWST. WEBB Space Telescope,

doi
[2014]

Accessed: 2024-11-28

URL https://science.nasa.gov/missions/hubble/ a-look-at-the-numbers-as-nasas-hubble-space-telescope-enters-its-25th-year/ . Accessed: 2024-11-28. NASA. Audit of the nancy grace roman space telescope project, Jul

2024
[2015]

doi: 10.1117/1.JATIS.1.1.014003. C. Rockosi et al. The low-resolution imaging spectrograph red channel CCD upgrade: fully depleted, high-resistivity CCDs for Keck. In Ian S. McLean, Suzanne K. Ramsay, and Hideki Takami (eds.),Ground-based and Airborne Instrumentation for Astro...

doi
[2017]

International networking in support of extremely large astronomical data-centric operations

Jeronimo Bezerra, Vinicius Arcanjo, Julio Ibarra, Jeff Kantor, Ron Lambert, Matt Kollross, Albert Astudillo, Shahram Sobhani, Sandra Jaque, Donald Petravick, Heidi Morgan, and Luiz Lopez. International networking in support of extremely large astronomical data-centric operatio...

2017
[2018]

Jonathan P

doi: 10.3390/galaxies6040120. Jonathan P. Gardner et al. The James Webb Space Telescope Mission.PASP, 135(1048):068001, June

doi
[2019]

URL https://link

doi: 10.1103/RevModPhys.91.045002. URL https://link. aps.org/doi/10.1103/RevModPhys.91.045002. CCSDS. Lossless Multispectral and Hyperspectral Image Compression,

doi
[2021]

Lucas Hayne, John Clyne, and Shaomeng Li

doi: 10.3847/2041-8213/abf2c7. Lucas Hayne, John Clyne, and Shaomeng Li. Using neural networks for two dimensional scientific data compression. In2021 IEEE International Conference on Big Data (Big Data), pp. 2956–2965. IEEE,

doi
[2022]

Accessed: 2024-11-28

URL https: //public.ccsds.org/Pubs/120x2g2.pdf. Accessed: 2024-11-28. CCSDS. Consultative committee for space data systems,

2024
[2023]

doi: 10.1088/1538-3873/acd1b5. Google. Neural Mapping,

doi
[2024]

org/default.aspx

URL https://public.ccsds. org/default.aspx. Accessed: 2024-11-28. Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers.arXiv preprint arXiv:1904.10509,

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

cosmic-ray [CR] hits

A.3.3 SBI-16-2D (HUBBLE) This dataset is based on data from the Wide Field Channel (WFC) instrument of the Advanced Camera for Surveys (ACS) onboard the Hubble Space Telescope (HST). Unlike the GBI-16-2D dataset, all observations in SBI-16-2D are obtained in space and with the same bandpass filter (F606W), providing a more uniform point-spread function ac...

work page

[2] [2]

Borucki, David Koch, Gibor Basri, Natalie Batalha, Timothy Brown, Douglas Caldwell, John Caldwell, Jørgen Christensen-Dalsgaard, William D

William J. Borucki, David Koch, Gibor Basri, Natalie Batalha, Timothy Brown, Douglas Caldwell, John Caldwell, Jørgen Christensen-Dalsgaard, William D. Cochran, Edna DeV ore, Edward W. Dunham, Andrea K. Dupree, Thomas N. Gautier, John C. Geary, Ronald Gilliland, Alan Gould, Steve B. Howell, Jon M. Jenkins, Yoji Kondo, David W. Latham, Geoffrey W. Marcy, Sø...

work page 2025

[3] [3]

reversible color transforms

and then randomly sampled fields and determined the center of the field sky positions. For each such field, we queried the SDSS Stripe 82 database for other images (regardless of quality) that overlap the field center and downloaded those images. Since field images from different runs are not aligned in RA, we also downloaded the two fields immediately ad...

work page 2023

[4] [5]

R” and “G

• To implement approach (1), we make the following modifications to support 16-bit input directly: ForIDF, we change the input scaling coefficient from28 to 216, so that it models the set Z/65536 (instead of Z/256). ForL3CandPixelCNN++, which both use a discretized logistic mixture likelihood model (Salimans et al., 2017), we increase the number of bins f...

work page 2019

[5] [6]

exascale

24 Published as a conference paper at ICLR 2025 Model / Dataset L3C-OpenImages (zero-padded) L3C-ImageNet64 (zero-padded) L3C-OpenImages (duplicated) L3C-ImageNet64 (duplicated) IDF-Cifar10 (zero-padded) IDF-Cifar10 (duplicated) LCO 1.78 2.16 2.04 2.05 1.72 1.84 Keck 1.26 1.44 1.27 1.38 0.66 0.59 Hubble 1.77 2.18 1.88 1.90 0.59 1.08 JWST 1.08 1.19 1.06 1....

work page 2025

[6] [7]

Neural data compression for physics plasma simulation

Jong Choi, Michael Churchill, Qian Gong, Seung-Hoe Ku, Jaemoon Lee, Anand Rangarajan, Sanjay Ranka, Dave Pugmire, CS Chang, and Scott Klasky. Neural data compression for physics plasma simulation. InNeural Compression: From Information Theory to Applications–Workshop@ ICLR 2021,

work page 2021

[7] [10]

Accessed: 2024-11-28

URL https://sites.research.google/ neural-mapping/. Accessed: 2024-11-28. Hannah Gulick, Jessica Lu, Steve Beckwith, Joshua Bloom, and Guy Nir. CubeSat for Rapid IR and Optical Surveys. In44th COSPAR Scientific Assembly. Held 16-24 July, volume 44, pp. 1985, July

work page 2024

[8] [12]

Jenkins and Jeb Dunnuck

12 Published as a conference paper at ICLR 2025 Jon M. Jenkins and Jeb Dunnuck. The little photometer that could: technical challenges and science results from the Kepler Mission. In Howard A. MacEwen and James B. Breckinridge (eds.),UV/Optical/IR Space Telescopes and Instruments: Innovative Technologies and Concepts V, volume 8146, pp. 814602. Internatio...

work page 2025