Transformer-Based Neural Network for Transient Detection without Image Subtraction

Adi Inada; Federica Bianco; Masao Sako; Tatiana Acero-Cuellar

REVIEW 3 major objections 5 minor 21 references

Transformer-Based Neural Network for Transient Detection without Image Subtraction

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A transformer can spot real versus bogus supernovae without image subtraction.

desk verdict Solid empirical paper on transformer-based real/bogus classification without difference images; headline numbers are credible, but the off-center benchmark is filtered by DIA detectability, which narrows the 'without subtraction' claim. read the letter →

arxiv 2508.16844 v1 pith:KADZQBWZ submitted 2025-08-22 cs.CV astro-ph.IM

Adi Inada , Masao Sako , Tatiana Acero-Cuellar , Federica Bianco This is my paper

classification cs.CVastro-ph.IM

keywords real-bogusclassificationtransientdetectiontransformerlocalizedattentiondifferenceimageanalysissupernovaDarkEnergySurveysubtraction

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

This paper claims that a transformer-based neural network can classify real supernova detections versus artifacts almost as accurately from the search and template images alone as from the usual three-image set that includes the computationally expensive difference image. On the autoScan dataset of Dark Energy Survey stamps, the two-image network reaches 97.4% accuracy and ROC AUC 0.993; adding the difference image improves accuracy by only 0.4 percentage points, and that gap shrinks as training data grows. On a custom no-Diff dataset where sources are not required to be centered, the network still reaches 96.7% accuracy, suggesting the method tolerates transients appearing anywhere in the stamp. The practical stake is that large surveys could run real-bogus triage without producing difference images, saving substantial compute per CCD.

What carries the argument

The central mechanism is the localized attention module. It augments standard self-attention with a learnable distance-weighting term: the attention logits are multiplied elementwise by exp(αD), where D contains pairwise pixel distances and α is a trainable parameter; this lets the model decide how local the comparison should be. Combined with channel swapping between the search and template feature maps, it turns the network into a dedicated pixel-comparison engine, and the decoder blocks fuse the two streams into a single real/bogus prediction.

What would settle it

Retrain or evaluate the network on a version of the no-Diff dataset where every injected fake with search-image SNR>5 counts as positive, regardless of whether it appears in the difference image, and compare the fake recovery rate; a large drop would show the current result is conditioned on difference-image detectability rather than true off-center robustness.

Watch

Extended reading notes

Core claim

Real-bogus classification can be cast as a pixel-by-pixel comparison task between two registered images rather than as a single-image recognition task on a difference image. The network does this with localized attention: a modified attention operation re-weights the softmax attention matrix by a learned exponential of pairwise pixel distances, forcing the model to compare nearby pixels across the search and template stamps. Channel swaps at the end of each encoder block exchange half the feature channels between the two images so downstream blocks learn which spatial and channel features distinguish a genuine transient from a subtraction artifact. Six encoder-decoder stages feed an MLP that

Load-bearing premise

The off-center test keeps only fakes that were bright enough in the difference image, so the high accuracy does not prove the network can catch transients that difference imaging would miss.

Editorial extensions

If this is right

A survey could classify candidates without running difference imaging at all; the authors report DiffImg takes about 10 minutes per CCD for DES while their network processes 1024 stamps in under 1.4 seconds.
The utility of the difference image as an input diminishes with training-set size, so larger labeled datasets make the no-subtraction route increasingly attractive.
Performance stays near 97% even when the transient is not centered in the stamp, meaning source extraction from the search image alone can feed the classifier.
Because the network reads raw fits inputs, it can be inserted directly into existing alert pipelines without special preprocessing.

Reading between the lines

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

The same distance-weighted attention design could apply to any task that asks whether two registered images differ for physical reasons, such as change detection in satellite imagery or medical follow-up imaging; the paper does not test these domains.
A natural next test is to train on positives defined by search-image detections without any difference-image selection; that would show whether the network can catch transients that DIA itself would miss, which the current no-Diff benchmark does not measure.
If the difference image's marginal value keeps shrinking with data, an ensemble or larger no-Diff model might eventually match or beat the triplet network, making DIA only a verification step rather than a discovery step.
The channel-swap trick could be read as a learned alternative to subtraction: instead of computing pixel differences, the network learns which feature channels to exchange, so future architectures might derive explicit difference-like operators from the attention weights.

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

The paper does what it says: a transformer with distance-weighted attention classifies DES real/bogus stamps using search and template images only, hitting 97.4% accuracy on autoScan, essentially matching the 97.8% of the triplet network that also gets the difference image. That is a real jump over the 91.1% CNN baseline from Acero-Cuellar et al., and the five-run means with standard deviations are a good sign. The architecture is clearly described, the data and code are public, and the runtime comparison (under 1.4s per 1024 stamps versus ~10 min per CCD for DiffImg) is useful context for Rubin-scale pipelines.

What is genuinely new here is not the transformer itself but the specific localized attention mechanism with a learnable distance weight, plus the construction of the no-Diff dataset to test off-center sources. The no-Diff data is a real effort: they run SExtractor on search images, cut stamps around detections, and use a group-based split to avoid leakage. That is worth credit.

Now the soft spots. The stress-test note is right: the no-Diff positives are kept only for fakes with available autoScan scores and difference-image SNR > 3.5. So every positive in that benchmark is DIA-recoverable by construction. The 96.7% accuracy therefore shows the network can classify off-center candidates when those candidates are already known to be detectable in the difference image. It does not test the case that would actually justify skipping DIA in a real survey—transients that DIA would miss but that are visible in the search image. That is a selection-function limitation, not an arithmetic error, but it should be stated plainly.

Second, the claim that the performance gap narrows with dataset size (Figure 5) rests on a single growth curve with no error bars and only 10 epochs. That is too thin to support a strong "diminishing utility" conclusion. Third, there is no ablation showing the distance-weighted attention itself is what buys the improvement; a plain ViT or CMT baseline on the same data would make the architectural claim convincing. The braai comparison is a reimplementation, which is fine but should be flagged.

The central result—that classification without subtraction can match classification with subtraction—holds up. The paper is worth a serious referee, but the authors should fix the framing of the no-Diff experiment and add ablations before it is accepted. I would bring it to a reading group; it is a useful and honest empirical contribution to a practical problem.

Referee Report

3 major / 5 minor

Summary. The paper introduces a transformer-based real/bogus classifier for transient detection that operates on search and template images only, omitting the difference image that conventional pipelines require. The architecture combines localized attention with a learned distance weighting, channel swaps between the two input images, and decoder blocks feeding an MLP. On the DES autoScan benchmark the authors report 97.4 ± 0.2% accuracy and 0.993 ROC AUC with search+template inputs, close to the 97.8 ± 0.1% of a triplet network that also receives the difference image. On a custom no-Diff dataset built from search-image source extraction, they report 96.7 ± 0.1% accuracy and 0.993 AUC, and conclude that the difference image provides diminishing value with larger training sets and that the network is robust to off-center transients.

Significance. If the results hold, this work offers a practical route to bypass computationally expensive difference imaging in large surveys such as Rubin/LSST, and the architecture contribution is plausible. The paper has several concrete strengths: five-seed runs with standard errors, a group-based train/test split to avoid leakage, public code and data, a runtime comparison against DiffImg, and a same-condition comparison with a re-trained braai baseline. The main limitation is that the no-Diff benchmark's positive samples are selected by DIA-derived scores and SNR, so the 'without subtraction' claim is only demonstrated for DIA-detectable transients. This is a correctness-risk concern for the paper's strongest conclusion, not an arithmetic error.

major comments (3)

[§2.2, §4.2] The no-Diff positive set is defined as fakes with 'available autoScan scores and SNR greater than 3.5 from the difference image.' This conditions the benchmark on the very DIA detectability that the paper argues can be bypassed. The reported 96.7 ± 0.1% accuracy therefore measures classification of DIA-detectable transients at off-center positions, not the ability to recover transients that search-image source extraction would find but DIA would miss. Please either construct positives without any DIA-based reference (e.g., search-image-only detections with independent visual labels), report performance on fakes below the DIA SNR threshold, or explicitly re-scope the title and conclusions to 'DIA-detectable transients.'
[§4.1, Figure 5] The claim that the utility of the difference image 'diminishes' with training-set size rests on a single growth curve of accuracy differential with no error bars, no indication of seed-to-seed variation, and only 10 epochs of training. Because this claim appears in the abstract and conclusions, it is load-bearing. Please add multiple-seed statistics at each dataset size (and ideally a final-epoch or fixed-schedule comparison) or soften the claim to a tentative observation.
[§4.1, Table 1] The comparison with the CNN baseline of Acero-Cuellar et al. is confounded by dataset size: their 91.1% and 96.1% accuracies are from 100,000 samples, whereas the attention network uses 800,768. The same-condition braai baseline partially addresses this, but the claim that the method 'outperforms traditional approaches' would be substantially stronger if the Acero-Cuellar CNN were retrained on the same 800,768 samples (or the attention network were evaluated on 100,000). Please either add this controlled comparison or note the confound explicitly.

minor comments (5)

[§3] Typo: 'an multi-layer perceptron' should be 'a multi-layer perceptron'.
[Table 1] The 'T able 1' rendering is corrupted; fix the table environment so the title appears correctly.
[§4.2] The ROC AUC is reported as '0.993 ± 0.0'. A non-zero standard deviation should be reported to at least two significant digits, or the uncertainty should be stated as below the printed precision.
[Figure 5] The y-axis label 'T wo' appears to contain a typo; it should read 'Two Network Configurations'.
[§2.1] The comparison of input preprocessing approaches (linear, Gaussian, raw FITS) is described in prose only. Consider a small table or figure showing the performance differences, since the choice of raw FITS is part of the method.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the benchmark is empirical, and the only mild self-reference is an evaluation filter, not a fitted input.

full rationale

The paper is an empirical benchmark study. The central claims (97.4% accuracy on autoScan and 96.7% on no-Diff) are measured against externally supplied labels: DES autoScan labels from Goldstein et al. (2015) and injected fakes from Kessler et al. (2015). The transformer's output is not constructed from its inputs by definition; it is trained and tested on held-out image pairs. The no-Diff dataset construction in Section 2.2 does keep only positives with available autoScan scores and difference-image SNR>3.5, so the 'without subtraction' result is conditional on DIA-detectability; this is a selection-function limitation that limits generalization to transients DIA would miss, but it is not circular because the labels still come from injected fakes and the network never receives the autoScan score or difference image as an input. The citation of Acero-Cuellar et al. (2023), which shares two co-authors, is used only as a prior benchmark and motivation; it is not load-bearing for the new architecture or its measured performance. No equation or fitted parameter is reused as a prediction, and no uniqueness theorem or ansatz is imported from the authors' prior work. Therefore the derivation chain is self-contained: measured accuracy against external labels, with one evaluation caveat.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical entities are introduced. The only created artifacts are the no-Diff dataset and the network architecture, neither of which is a postulated physical entity.

free parameters (3)

Network weights and biases = millions of learned values, not enumerated
All accuracy claims are outcomes of fitting model parameters to the training sets; no final model file is provided.
Learnable attention localization weight alpha = learned, not reported
The attention formula softmax(QK^T/sqrt(d)) * exp(alpha D) includes alpha, which controls how local attention is and is trained; the fitted value is not given.
Dataset-selection thresholds = SNR>5 source extraction; SNR>3.5 diff-image cutoff; 20% subsample
Hand-chosen thresholds define the no-Diff dataset and therefore which positive and negative examples are used; central to the off-center robustness claim.

assumptions (3)

domain assumption Search and template stamps are aligned and photometrically comparable before input
Localized pixel-by-pixel attention assumes corresponding pixels refer to the same sky location; no registration step is described in Sections 2 and 3.
domain assumption autoScan labels are sufficiently reliable
Training and evaluation use autoScan ground truth, but the paper itself notes a non-negligible fraction of mislabeled samples in Section 4.1.
domain assumption Injected DiffImg fakes represent real transients
Positive labels in both datasets come from fake injections rather than confirmed supernovae; this is a standard benchmark assumption but limits generalization.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Transformer-Based Neural Network for Transient Detection without Image Subtraction." pith.science (2026). https://pith.science/paper/KADZQBWZ

@misc{pith2026250816844,
  author       = {Pith},
  title        = {Pith review of: Transformer-Based Neural Network for Transient Detection without Image Subtraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KADZQBWZ}},
  note         = {Machine review of arXiv:2508.16844}
}

read the original abstract

We introduce a transformer-based neural network for the accurate classification of real and bogus transient detections in astronomical images. This network advances beyond the conventional convolutional neural network (CNN) methods, widely used in image processing tasks, by adopting an architecture better suited for detailed pixel-by-pixel comparison. The architecture enables efficient analysis of search and template images only, thus removing the necessity for computationally-expensive difference imaging, while maintaining high performance. Our primary evaluation was conducted using the autoScan dataset from the Dark Energy Survey (DES), where the network achieved a classification accuracy of 97.4% and diminishing performance utility for difference image as the size of the training set grew. Further experiments with DES data confirmed that the network can operate at a similar level even when the input images are not centered on the supernova candidate. These findings highlight the network's effectiveness in enhancing both accuracy and efficiency of supernova detection in large-scale astronomical surveys.

Figures

Figures reproduced from arXiv: 2508.16844 by the authors.

**Figure 1.** Stamps from no-Diff dataset. Positive (label = 1) and negative (label = 0) examples are shown. From the top row to the bottom row we show bad subtraction negative label, off the center positive label, clean subtraction negative label and cleanly subtracted positive label, respectively. Similar to the autoScan dataset, we used raw inputs to train our neural network and split the dataset into training and test subset … view at source ↗

**Figure 2.** Network architecture of real-bogus classifier used in this work. (Above) Input images are passed through six localized attention blocks with shared weights to extract salient features. Three distance weighted localized attention steps are performed in each block. The features are forwarded to decoder blocks and MLP layer for binary prediction. (Below) Inner working of localized attention modules. In our work we use … view at source ↗

**Figure 3.** ROC curve for the autoScan dataset with and without using difference image. The network trained on the entire dataset with a 9:1 train-test split. 19 20 21 22 23 24 25 26 Magnitude 0.80 0.85 0.90 0.95 1.00 Fake Recovery Rate Fake Recovery Rate vs Magnitude 0.0 0.2 0.4 0.6 0.8 1.0 Flux Ratio 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 Fake Recovery Rate Fake Recovery Rate vs Flux Ratio [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Fake recovery rate (true positive rate) analysis on the autoScan test data. (Left) Fake recovery rate as a function of magnitude. (Right) Fake recovery rate as a function of flux ratio. The flux ratio is defined as clip 0, 1, fd ft+fd where fd and ft are fluxes of …

**Figure 5.** Figure 5: The difference in accuracy between search-template-diff network and search-template network as a function of dataset size for the autoScan dataset after 10 epochs of training. We summarize our network performance and compare them to other published results in [PITH_FU…

**Figure 6.** Figure 6: ROC Curve for no-Diff test dataset [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 7.** Figure 7: Fake recovery rate (true positive rate) analysis on the no-Diff dataset. (Left) Fake recovery rate as a function of SN magnitude. Bins with at least 25 samples are shown. (Right) Fake recovery rate as a function of flux ratio. The flux ratio is defined as clip 0, 1, …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 16 canonical work pages

[1]

2023, The Astronomical Journal, 166, 115

Acero-Cuellar, T., Bianco, F., Dobler, G., et al. 2023, The Astronomical Journal, 166, 115

work page 2023
[2]

Alard, C., & Lupton, R. H. 1998, The Astrophysical Journal, 503, 325

work page 1998
[3]

1989, A&AS, 81, 421

Barbon, R., Cappellaro, E., & Turatto, M. 1989, A&AS, 81, 421

work page 1989
[4]

S., Richards, J

Bloom, J. S., Richards, J. W., Nugent, P. E., et al. 2012, Publications of the Astronomical Society of the Pacific, 124, 1175

work page 2012
[5]

Bramich, D. M. 2008, Monthly Notices of the Royal Astronomical Society: Letters, 386, L77

work page 2008
[6]

W., Poznanski, D., et al

Brink, H., Richards, J. W., Poznanski, D., et al. 2013, Monthly Notices of the Royal Astronomical Society, 435, 1047

work page 2013
[7]

2020, End-to-End Object Detection with Transformers, computer Vision and Pattern Recognition, arXiv:2005.12872

Carion, N., Massa, F., Synnaeve, G., et al. 2020, End-to-End Object Detection with Transformers, computer Vision and Pattern Recognition, arXiv:2005.12872

arXiv 2020
[8]

Chen, Z., Zhou, W., Sun, G., et al. 2023, TransientViT: A novel CNN - Vision Transformer hybrid real/bogus transient classifier for the Kilodegree Automatic Transient Survey, instrumentation and Methods for Astrophysics, arXiv:2309.09937 Dan Hendrycks, K. G. 2016, arXiv:1606.08415v5, doi:https://doi.org/10.48550/arXiv.1606.08415

work page Pith review arXiv doi:10.48550/arx 2023

Show all 21 references

[9]

2020, An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, computer Vision and Pattern Recognition, arXiv:2010.11929

Dosovitskiy, A., Beyer, L., Kolesnikov, A., et al. 2020, An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, computer Vision and Pattern Recognition, arXiv:2010.11929

2020 arXiv
[10]

A., Mahabal, A., Masci, F

Duev, D. A., Mahabal, A., Masci, F. J., et al. 2019, Monthly Notices of the Royal Astronomical Society, 489, 3582

2019
[11]

A., D’Andrea, C

Goldstein, D. A., D’Andrea, C. B., Fischer, J. A., et al. 2015, The Astronomical Journal, 150, 82

2015
[12]

2017, ApJ, 835, 64

Margutti, R. 2017, ApJ, 835, 64

2017
[13]

2022, in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 12165–12175

Guo, J., Han, K., Wu, H., et al. 2022, in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 12165–12175

2022
[14]

2022, The Astrophysical Journal, 936, 157

Hu, L., Wang, L., Chen, X., & Yang, J. 2022, The Astrophysical Journal, 936, 157

2022
[15]

2015, AJ, 150, 172

Kessler, R., Marriner, J., Childress, M., et al. 2015, AJ, 150, 172

2015
[16]

2019, Publications of the Astronomical Society of the Pacific, 131, 038002

Mahabal, A., Rebbapragada, U., Walters, R., et al. 2019, Publications of the Astronomical Society of the Pacific, 131, 038002

2019
[17]

2016, Publications of the Astronomical Society of Japan, 68, 104

Morii, M., Ikeda, S., Tominaga, N., et al. 2016, Publications of the Astronomical Society of Japan, 68, 104

2016
[18]

1999, The Astrophysical Journal, 517, 565

Perlmutter, S., Aldering, G., Goldhaber, G., & et al. 1999, The Astrophysical Journal, 517, 565

1999
[19]

G., Filippenko, A

Riess, A. G., Filippenko, A. V., Challis, P., & et al. 1998, The Astronomical Journal, 116, 1009 The Dark Energy Survey Collaboration. 2005, The Dark Energy Survey: More than Dark Energy – An Overview, arXiv e-prints, arXiv:astro-ph/0510346

1998 arXiv
[20]

2017, in Advances in Neural Information Processing Systems, ed

Vaswani, A., Shazeer, N., Parmar, N., et al. 2017, in Advances in Neural Information Processing Systems, ed. I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, & R. Garnett, Vol. 30 (Curran Associates, Inc.)

2017
[21]

E., Smartt, S

Wright, D. E., Smartt, S. J., Smith, K. W., et al. 2015, Monthly Notices of the Royal Astronomical Society, 449, 451

2015

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

2023, The Astronomical Journal, 166, 115

Acero-Cuellar, T., Bianco, F., Dobler, G., et al. 2023, The Astronomical Journal, 166, 115

work page 2023

[2] [2]

Alard, C., & Lupton, R. H. 1998, The Astrophysical Journal, 503, 325

work page 1998

[3] [3]

1989, A&AS, 81, 421

Barbon, R., Cappellaro, E., & Turatto, M. 1989, A&AS, 81, 421

work page 1989

[4] [4]

S., Richards, J

Bloom, J. S., Richards, J. W., Nugent, P. E., et al. 2012, Publications of the Astronomical Society of the Pacific, 124, 1175

work page 2012

[5] [5]

Bramich, D. M. 2008, Monthly Notices of the Royal Astronomical Society: Letters, 386, L77

work page 2008

[6] [6]

W., Poznanski, D., et al

Brink, H., Richards, J. W., Poznanski, D., et al. 2013, Monthly Notices of the Royal Astronomical Society, 435, 1047

work page 2013

[7] [7]

2020, End-to-End Object Detection with Transformers, computer Vision and Pattern Recognition, arXiv:2005.12872

Carion, N., Massa, F., Synnaeve, G., et al. 2020, End-to-End Object Detection with Transformers, computer Vision and Pattern Recognition, arXiv:2005.12872

arXiv 2020

[8] [8]

Chen, Z., Zhou, W., Sun, G., et al. 2023, TransientViT: A novel CNN - Vision Transformer hybrid real/bogus transient classifier for the Kilodegree Automatic Transient Survey, instrumentation and Methods for Astrophysics, arXiv:2309.09937 Dan Hendrycks, K. G. 2016, arXiv:1606.08415v5, doi:https://doi.org/10.48550/arXiv.1606.08415

work page Pith review arXiv doi:10.48550/arx 2023