Pith. sign in

REVIEW 4 major objections 6 minor 62 references

BCDDM: Branch-Corrected Denoising Diffusion Model for Black Hole Image Generation

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read This paper proposes BCDDM, a diffusion model that generates black hole images from seven accretion-flow parameters, and shows that adding these synthetic images to the training set improves a parameter regression network.

desk verdict A useful first application of diffusion to black hole image generation, but the augmentation claim lacks a size-matched control and may leak the test set through validation. read the letter →

arxiv 2502.08528 v3 pith:JB4DEETM submitted 2025-02-12 astro-ph.GA cs.CV

classification astro-ph.GAcs.CV
keywords denoisingdiffusionmodelblackholeimagegenerationgeneralrelativisticraytracingradiativelyinefficientaccretionflowdataaugmentationparameterregressionbranchcorrection
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

The paper proposes Branch-Corrected Denoising Diffusion Model (BCDDM), a conditional diffusion model that generates 230 GHz black hole images directly from seven physical parameters of a radiatively inefficient accretion flow: spin, mass, electron temperature, disk thickness, Keplerian factor, position angle, and flow direction. The model adds a parameter-correction branch to a U-Net and trains with a weighted sum of noise-prediction and label-consistency losses, so that generated images are tied to the input parameters rather than just looking plausible. On a training set of 2,157 general-relativistic ray-traced images, BCDDM reconstructs held-out images with high structural similarity and predicts parameters that closely match the ground truth. When the synthetic images are mixed with real ones and used to train a ResNet50 regressor, most parameters show higher $R^2$ scores than training on real images alone. This is a fast surrogate for ray tracing—about 5.25 seconds per image—and a new data-augmentation route for black-hole parameter estimation.

What carries the argument

The load-bearing mechanism is the Branch-Corrected U-Net: a standard diffusion U-Net whose bottleneck is augmented with a parameter-correction branch that maps the latent feature vector to the seven physical parameters. The model is trained to minimize a weighted mixed loss $\mathcal{L}=\lambda_1\mathcal{L}_{\mathrm{noise}}+\lambda_2\mathcal{L}_{\mathrm{label}}$ (with $\lambda_1=0.95$, $\lambda_2=0.05$), where $\mathcal{L}_{\mathrm{noise}}$ is the standard denoising objective and $\mathcal{L}_{\mathrm{label}}$ forces the latent representation to encode the parameters. During sampling, the time step and parameter vector condition the denoising of $x_T$ back to $x_0$, yielding an image with the requested physical properties.

What would settle it

Run the paper's regression experiment on a fresh, independently generated GRRT test set of several hundred images drawn uniformly from the Table 1 ranges; if the mixed-dataset $R^2$ advantage over real-only training does not reproduce, or if the 20 μas-blurred disk-thickness $R^2$ remains negative (the paper already reports −0.3325), the claim that synthetic augmentation reliably improves parameter prediction would be refuted.

Watch

Extended reading notes

Core claim

BCDDM's central claim is that a diffusion model can learn the mapping from the seven RIAF (radiatively inefficient accretion flow) parameters to the simulated image, and that the learned mapping is accurate enough to serve as a data generator. The branch-correction architecture forces the latent representation at the U-Net bottleneck to predict the input labels, so the model optimizes both the denoising error and the parameter-consistency error. The paper reports SSIM values of 0.877–0.975 on six reconstructed test images, with NRMSE sometimes high because of spatial misalignment between sampled and target images; the parameter branch returns values close to the ground truth, with small deviations in spin $a$ and electron temperature $T_e$. For the regression evaluation, mixing BCDDM-generated images with real ones raises the test $R^2$ from 0.9206 to 0.9645 for $a$, from 0.9671 to 0.9960 for $T_e$, and from 0.9012 to 0.9602 for position angle, while the binary flow-direction accuracy rises from 92.19% to 94.27%. The paper also reports that outside the training parameter range the model generalizes poorly, so the mapping is reliable mainly inside the sampled box.

Load-bearing premise

The whole method assumes that 2,157 simulated images are enough for the model to learn a smooth, reliable mapping between the seven physical parameters and the image, so that images it generates fall on the same distribution as true ray-traced images.

Editorial extensions

If this is right

  • A regressor trained on real plus synthetic images achieves higher $R^2$ for most parameters than one trained on real images alone, so BCDDM can augment small GRRT datasets without breaking their physical statistics.
  • Because generation takes about 5.25 seconds per image on one GPU, a researcher can expand a 2,157-image training set by thousands of samples at a fraction of the ray-tracing cost.
  • The parameter-correction branch itself acts as a fast estimator of black hole parameters from images, providing a second route to parameter inference within the same model.
  • The method is not tied to the RIAF model; the authors state it can be retrained on other accretion models, and with multi-channel inputs it could be extended to polarized images.
  • Because the model fails outside the parameter ranges it was trained on, any practical use for survey-level parameter estimation would need a substantially wider training set than the 2,157-image RIAF dataset.

Reading between the lines

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

  • Editorial inference: If the conditional diffusion mapping is smooth inside the training box, the same architecture could be used as an amortized surrogate inside a Bayesian likelihood evaluation, replacing on-the-fly GRRT calls during MCMC sampling; the paper does not test this.
  • Editorial inference: The sharp drop in disk-thickness $R^2$ at 20 μas blur (from 0.9717 to −0.0683 for real data) suggests that no amount of synthetic augmentation can restore a feature the instrument cannot resolve; accurate estimation of $h_{\rm disk}$ from images at this resolution may require an explicit blurring or multi-epoch model.
  • Editorial inference: The larger $R^2$ gains for position angle (+0.0590) and the flow-direction accuracy gain (92.19% to 94.27%), versus essentially no gain for mass (≈0.0000), suggest that augmentation helps most for parameters with subtle or orientation-dependent image signatures; a direct test would compare per-parameter learning curves on synthetic-only versus real-only data.
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.

Referee Report

4 major / 6 minor

Summary. The paper presents BCDDM, a conditional denoising diffusion model for generating black hole accretion-flow images from seven physical parameters (spin, mass, electron temperature, disk thickness, Keplerian factor, position angle, and flow direction). The model is trained on 2,157 GRRT (ipole) images of a RIAF model, with a novel branch-correction architecture and a mixed loss combining noise prediction and label prediction. The authors report NRMSE/SSIM reconstruction metrics and evaluate the usefulness of the generated images as data augmentation for a ResNet50 regressor that predicts black hole parameters, comparing datasets of real, fake, and mixed images.

Significance. BCDDM is, to the authors' knowledge, the first diffusion-based generative model for black hole images, and the public availability of code and data is a strength. The reported generation speed of 5.25 s per image, though slower than direct GRRT surrogates, offers a concrete speed-up over computational GRRT and the approach is potentially extensible to other accretion models and polarization. However, the central claims of high-fidelity generation and augmentation benefit are not yet fully supported by the evidence: the NRMSE values are large, the data augmentation comparisons lack a size-matched control, and the label loss introduces a circularity in demonstrating parameter-image consistency. With additional experimental controls and statistical rigor, the method could become a useful tool.

major comments (4)
  1. [Section 3.2, Table 3] The comparison between RLDs (1,725 training images) and MXDs (3,450 training images, 1,725 real + 1,725 BCDDM) does not control for dataset size. The observed R2 increases (e.g., a: 0.9206 to 0.9645; PA: 0.9012 to 0.9602) could be driven by the additional training data rather than the physical fidelity of BCDDM images. A size-matched control, such as training on 3,450 real GRRT images or on the RLDs with a classical augmentation (noise, small shifts, or duplicated real images), is required to support the claim that BCDDM augmentation provides unique information. Furthermore, Table 4 shows that the MXD performance for hdisk at 20 µas blurring is worse than RLDs (-0.3325 vs -0.0683), which is inconsistent with a robust augmentation benefit.
  2. [Section 2.4, Eq. (11)] The label loss L_label explicitly optimizes the branch to predict the conditioning parameters from the intermediate feature representation of the generated image. This enforces a statistical association between generated images and parameters by construction, so the strong parameter-prediction results in Figure 5 and the R2 improvements in Table 3 are not independent evidence that the images encode physically correct features. Since the regressor is evaluated on real test images, the improvement in MXDs could reflect the regressor exploiting label-specific cues present in the synthetic images; the paper should demonstrate that the augmentation benefit persists when the label branch is ablated or when the comparison is size-matched.
  3. [Section 3.1, Figure 5] The claim that BCDDM generates “clear and high-quality” black hole images is not consistent with the reported NRMSE values, which range from 0.072 to 1.037 across the six test images, with panels (b) and (d) at 0.897 and 1.037. The authors attribute these discrepancies to spatial alignment and brightness instability, but this undercuts the utility of the model as a pixel-accurate surrogate. Additionally, Figure 6 shows that outside the training parameter ranges the model produces high NRMSE and low SSIM (e.g., panel f: NRMSE=1.349, SSIM=0.165), confirming that the surrogate is only valid within the narrow training distribution.
  4. [Section 3.2, Tables 3 and 4] No error bars, confidence intervals, or repeated-seed experiments are reported for any R2 value. With a test set of only 216 images, the differences between RLDs and MXDs are subject to sampling noise; for example, the hdisk R2 of 0.9841 versus 0.9893 in Table 3 is small relative to the likely variance. The claim of “significant improvements” requires either multiple training runs with reported variability or a statistical significance test.
minor comments (6)
  1. [Section 2.2] The heading 'black hole image dateset' should be corrected to 'dataset'.
  2. [Throughout] The parameter hdisk is sometimes denoted 'h' (e.g., Section 3.1, Figure 6); please use consistent notation.
  3. [Section 2.4, Figures 2 and 3] The architecture is hard to parse; the feeding of the conditioning label at training and sampling, and the role of the predicted label at inference, should be clarified.
  4. [Section 2.3] The linear noise schedule (β_t from 1e-4 to 0.02) is chosen without motivation; a discussion or comparison with alternative schedules would strengthen the paper.
  5. [Section 3.2] For the binary parameter Fdir, R2 is not an appropriate regression metric; the confusion matrix in Figure 8 is more informative, and the regression formulation for a binary variable should be explained.
  6. [References] The citation 'Wan & Ohtani 2000' for Eq. (11) appears unrelated to the label loss; please verify the reference.

Circularity Check

1 steps flagged · score 6.0 of 10

Label-loss-trained parameter branch makes the 'strong correlation' evidence self-confirming; SSIM/NRMSE and FKDs checks provide non-circular grounding.

  1. fitted input called prediction [Section 2.4, Eq. (11), used as evidence in Section 3.1 and Fig. 5]
    "Meanwhile, the label loss, Llabel = Ex0,l h (f (k) − l)2 i , enforces alignment between generated images and their target physical parameters. ... The physical parameters predicted by the parameter correction branch closely match the original values ... This confirms that the parameter correction branch effectively learns the mapping between images and physical parameters, demonstrating BCDDM’s sensitivity to parameter variations."

    The quoted 'prediction' is made by the parameter-correction branch that is explicitly trained, via L_label, to output the ground-truth physical parameters from image features, and during sampling the same label is also injected as conditioning ('we encode both the step t and the image’s label, embedding this information into the noisy image xt'). The near-match between the branch output and the input label is therefore a consequence of the training objective and the conditioning channel, not an independent measurement that the image content alone encodes the parameters. Using this agreement to 'confirm' that BCDDM-generated image features correspond to the parameters is a fitted output being presented as its own validation.

full rationale

The derivation is not globally circular: BCDDM is trained on 1,941 GRRT-rendered RIAF images and evaluated on 216 held-out images, and the SSIM/NRMSE comparisons against real GRRT images are independent, non-circular evidence that the generated images are structurally similar. The FKDs-only regressor underperforming RLDs on several parameters is also an honest external check. However, one load-bearing supporting claim is circular: the 'strong correlation between generated images and their physical parameters' is validated with the parameter-correction branch, which is trained by L_label to output exactly those labels while the labels are also injected as conditioning inputs, so the agreement is enforced by construction rather than independently established. The data-augmentation comparison (RLDs with 1,725 training images vs MXDs with 3,450 training images) is additionally confounded by dataset size; this is an experimental-control problem rather than a formal circularity, and it should be addressed with a size-matched control and repeated-seed error bars before the augmentation benefit is attributed to BCDDM fidelity. On balance, the central image-generation claim has independent content, but the parameter-correlation evidence and the 'proves in reverse' inference are partially self-confirming, giving a partial-circularity score of 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a set of standard generative-modeling math, a physics simulation assumption (RIAF, ipole) taken from prior literature, and a hand-chosen set of hyperparameters. The main load-bearing choices are the loss weights (lambda2 must be large enough to enforce label consistency) and the representativeness of the small, fixed-geometry dataset. No new physical entities are introduced.

free parameters (3)
  • Loss weights lambda1, lambda2 = 0.95, 0.05
    Chosen by hand (Section 3.1) to balance noise loss and label loss; the central claim depends on the label loss being effective, so these weights are load-bearing hyperparameters.
  • Diffusion noise schedule beta_t and T = beta from 1e-4 to 0.02, T=1000
    Standard DDPM schedule but still a hand-chosen configuration; the generation quality and sampling speed depend on it.
  • Training epochs and checkpoint selection = 8000 epochs, model saved at epoch 5000
    The paper uses the epoch-5000 checkpoint for all generation results; this is a post-hoc selection based on validation loss and affects the reported image quality.
assumptions (4)
  • domain assumption The RIAF model with thermal synchrotron emission is a valid description of the M87* accretion flow at 230 GHz.
    Section 2.1 adopts this model from Narayan et al. 1997 and Yuan & Narayan 2014. If the RIAF model is not representative of real EHT sources, the generated images and the parameter regression results do not transfer to actual observations.
  • domain assumption The ipole GRRT code produces ground-truth images for the target physical parameters.
    Section 2.1 uses ipole (Mosibrodzka & Gammie 2018) as the simulator. The entire training set and all real-image test labels come from this code; no external validation against EHT images is performed.
  • standard math The standard DDPM forward and reverse process equations (Eqs. 4-9) are correct.
    The paper uses the standard Ho et al. 2020 formulation without derivation; this is accepted background.
  • ad hoc to paper Adding the label prediction branch and the L_label loss enforces physical consistency between generated images and conditioning parameters.
    This is the paper's key architectural assumption (Section 2.4). The efficacy is asserted and tested only indirectly through SSIM/NRMSE and regressor performance; it is not guaranteed by the diffusion framework.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BCDDM: Branch-Corrected Denoising Diffusion Model for Black Hole Image Generation." pith.science (2026). https://pith.science/paper/JB4DEETM

@misc{pith2026250208528,
  author       = {Pith},
  title        = {Pith review of: BCDDM: Branch-Corrected Denoising Diffusion Model for Black Hole Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JB4DEETM}},
  note         = {Machine review of arXiv:2502.08528}
}
read the original abstract

The properties of black holes and accretion flows can be inferred by fitting Event Horizon Telescope (EHT) data to simulated images generated through general relativistic ray tracing (GRRT). However, due to the computationally intensive nature of GRRT, the efficiency of generating specific radiation flux images needs to be improved. This paper introduces the Branch Correction Denoising Diffusion Model (BCDDM), a deep learning framework that synthesizes black hole images directly from physical parameters. The model incorporates a branch correction mechanism and a weighted mixed loss function to enhance accuracy and stability. We have constructed a dataset of 2,157 GRRT-simulated images for training the BCDDM, which spans seven key physical parameters of the radiatively inefficient accretion flow (RIAF) model. Our experiments show a strong correlation between the generated images and their physical parameters. By enhancing the GRRT dataset with BCDDM-generated images and using ResNet50 for parameter regression, we achieve significant improvements in parameter prediction performance. BCDDM offers a novel approach to reducing the computational costs of black hole image generation, providing a faster and more efficient pathway for dataset augmentation, parameter estimation, and model fitting.

Figures

Figures reproduced from arXiv: 2502.08528 by the authors.

Figure 1
Figure 1. The diffusion process of black hole images. We output the image every 100 steps for a total of 1000 times to observe the changes in the diffusion process of the image. The diffusion model initiates with a forward diffusion process applied to the black hole image, wherein Gaus￾sian noise is incrementally introduced according to a predefined noise schedule. This process systematically corrupts the input image over T s… view at source ↗
Figure 2
Figure 2. BCDDM architecture for black hole image generation. The diffusion process and denoising process utilize the same U-Net model, with detailed architecture illustrated in [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Encoding and Decoding architecture embedded in BCDDM. The trainable components consist solely of the input encoded label vector and black hole image, along with the output predicted parameters and noise. To optimize the model, we design a composite loss function combining noise prediction loss Lnoise and parameter consistency loss Llabel (Wan & Ohtani 2000). The noise loss, defined as Lnoise = Et,x0,ϵ h (ϵ − ϵθ(xt, … view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Training loss and validation loss during the training process. The right panel displays two independent black hole images generated by BCDDM at epochs 100, 1000, 2500, and 5000, illustrating the performance progression of the generative model. loss plateaued after 5000…
Figure 5
Figure 5. Figure 5: Comparison of original and reconstructed black hole images in the additional test set. The figure presents six pairs of GRRT-simulated original black hole images (top row) alongside their BCDDM-reconstructed counterparts (bottom row). The parameters displayed above the…
Figure 6
Figure 6. Figure 6: Comparison of the original and reconstructed black hole images, similar to [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Workflow of Constructing Datasets from Multiple Sources and Evaluating Regression Model Performance. This figure outlines the process of constructing datasets from diverse sources and evaluating the performance of a re￾gression model. It highlights two primary data sou…
Figure 8
Figure 8. Figure 8: The binary confusion matrix of Fdir. The datasets corresponding to the three images are RLDs (left), FKDs (middle), and MXDs (right). The classification accuracies of three training sets for fluid-direction are 92.19%, 89.58% and 94.27%. At the same time, as shown in …
Figure 9
Figure 9. Figure 9: The fitting results of RLDs and MXDs on the test set samples.The prediction error of the test set after training with RLDs for 5000 epochs (left) and the prediction error of the test set after training with MXDs for 5000 epochs (right) [PITH_FULL_IMAGE:figures/full_fi…
Figure 10
Figure 10. Figure 10: The fitting results of RLDs and MXDs on the test set samples. The prediction error of the test set after training with RLDs for 5000 epochs (left) and the prediction error of the test set after training with MXDs for 5000 epochs (right). images while preserving the ch…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 26 canonical work pages

  1. [1]

    A., & Nazim Uddin, M

    Asgar, S., Sahil, A. A., & Nazim Uddin, M. 2023, in 2023 26th International Conference on Computer and Information Technology (ICCIT), 1–5, doi: 10.1109/ICCIT60459.2023.10441276

  2. [2]

    Bardeen, J. M. 1973, in Black Holes (Les Astres Occlus), ed. C. Dewitt & B. S. Dewitt, 215–239

  3. [3]

    E., Fish, V

    Broderick, A. E., Fish, V. L., Doeleman, S. S., & Loeb, A. 2011a, ApJ, 735, 110, doi: 10.1088/0004-637X/735/2/110 —. 2011b, ApJ, 738, 38, doi: 10.1088/0004-637X/738/1/38

  4. [4]

    E., Gold, R., Karami, M., Preciado-L´ opez, J

    Broderick, A. E., Gold, R., Karami, M., Preciado-L´ opez, J. A., et al. 2020, ApJ, 897, 139, doi: 10.3847/1538-4357/ab91a4

  5. [5]

    E., Fish, V

    Broderick, A. E., Fish, V. L., Johnson, M. D., et al. 2016, ApJ, 820, 137, doi: 10.3847/0004-637X/820/2/137

  6. [6]

    D., & Lupsasca, A

    Chael, A., Johnson, M. D., & Lupsasca, A. 2021, ApJ, 918, 6, doi: 10.3847/1538-4357/ac09ee

  7. [7]

    2018, ApJ, 867, 59, doi: 10.3847/1538-4357/aadfe5

    Chan, C.-k., Medeiros, L., ¨Ozel, F., & Psaltis, D. 2018, ApJ, 867, 59, doi: 10.3847/1538-4357/aadfe5

  8. [8]

    2024, in IAU General Assembly, 1132, doi: 10.48550/arXiv.2405.04749

    Chang, D., Tiede, P., Johnson, M., & Palumbo, D. 2024, in IAU General Assembly, 1132, doi: 10.48550/arXiv.2405.04749

Show all 62 references
  1. [9]

    2023, Science China Physics, Mechanics, and Astronomy, 66, 260401, doi: 10.1007/s11433-022-2059-5 Event Horizon Telescope Collaboration, Akiyama, K.,

    Chen, S., Jing, J., Qian, W.-L., & Wang, B. 2023, Science China Physics, Mechanics, and Astronomy, 66, 260401, doi: 10.1007/s11433-022-2059-5 Event Horizon Telescope Collaboration, Akiyama, K.,

  2. [10]

    2019a, ApJL, 875, L1, doi: 10.3847/2041-8213/ab0ec7 —

    Alberdi, A., Alef, W., et al. 2019a, ApJL, 875, L1, doi: 10.3847/2041-8213/ab0ec7 —. 2019b, ApJL, 875, L5, doi: 10.3847/2041-8213/ab0f43 —. 2022a, ApJL, 930, L12, doi: 10.3847/2041-8213/ac6674 —. 2022b, ApJL, 930, L17, doi: 10.3847/2041-8213/ac6756 —. 2022c, ApJL, 930, L16, do...

  3. [11]

    C., et al

    Algaba, J. C., et al. 2021a, ApJL, 910, L12, doi: 10.3847/2041-8213/abe71d —. 2021b, ApJL, 910, L13, doi: 10.3847/2041-8213/abe4de Event Horizon Telescope Collaboration, Akiyama, K.,

  4. [12]

    2024b, ApJL, 964, L26, doi: 10.3847/2041-8213/ad2df1 —

    Alberdi, A., et al. 2024b, ApJL, 964, L26, doi: 10.3847/2041-8213/ad2df1 —. 2024c, ApJL, 964, L25, doi: 10.3847/2041-8213/ad2df0

  5. [13]

    2000, ApJL, 528, L13, doi: 10.1086/312423

    Falcke, H., Melia, F., & Agol, E. 2000, ApJL, 528, L13, doi: 10.1086/312423

  6. [14]

    2021, in 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 142–151

    Fei, N., Gao, Y., Lu, Z., & Xiang, T. 2021, in 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 142–151

  7. [15]

    S., Palumbo, D

    Gelles, Z., Prather, B. S., Palumbo, D. C. M., Johnson, M. D., et al. 2021, ApJ, 912, 39, doi: 10.3847/1538-4357/abee13 19

  8. [16]

    J., Pouget-Abadie, J., Mirza, M., Xu, B., et al

    Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., et al. 2014, in Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 2, NIPS’14 (Cambridge, MA, USA: MIT Press), 2672–2680

  9. [17]

    E., Holz, D

    Gralla, S. E., Holz, D. E., & Wald, R. M. 2019, PhRvD, 100, 024018, doi: 10.1103/PhysRevD.100.024018

  10. [18]

    2016, in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770–778

    He, K., Zhang, X., Ren, S., & Sun, J. 2016, in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770–778

  11. [19]

    2020, Denoising Diffusion Probabilistic Models

    Ho, J., Jain, A., & Abbeel, P. 2020, Denoising Diffusion Probabilistic Models

  12. [20]

    1997, Neural computation, 9, 1735, doi: 10.1162/neco.1997.9.8.1735

    Hochreiter, S., & Schmidhuber, J. 1997, Neural computation, 9, 1735, doi: 10.1162/neco.1997.9.8.1735

  13. [21]

    2024, arXiv e-prints, arXiv:2409.07248, doi: 10.48550/arXiv.2409.07248

    Hou, Y., Huang, J., Mizuno, Y., Guo, M., et al. 2024, arXiv e-prints, arXiv:2409.07248, doi: 10.48550/arXiv.2409.07248

  14. [22]

    2024, JCAP, 2024, 054, doi: 10.1088/1475-7516/2024/11/054

    Huang, J., Zheng, L., Guo, M., & Chen, B. 2024, JCAP, 2024, 054, doi: 10.1088/1475-7516/2024/11/054

  15. [23]

    V., Narayan, R., & Abramowicz, M

    Igumenshchev, I. V., Narayan, R., & Abramowicz, M. A. 2003, ApJ, 592, 1042, doi: 10.1086/375769

  16. [24]

    D., Lupsasca, A., Strominger, A., Wong, G

    Johnson, M. D., Lupsasca, A., Strominger, A., Wong, G. N., et al. 2020, Science Advances, 6, eaaz1310, doi: 10.1126/sciadv.aaz1310

  17. [25]

    P., & Welling, M

    Kingma, D. P., & Welling, M. 2022, Auto-Encoding Variational Bayes. https://arxiv.org/abs/1312.6114

  18. [26]

    Krizhevsky, A., Sutskever, I., & Hinton, G. E. 2012, Communications of the ACM, 60, 84

  19. [27]

    2022, Science China

    Liu, X., Chen, S., & Jing, J. 2022, Science China

  20. [28]

    Physics, Mechanics, and Astronomy, 65, 120411, doi: 10.1007/s11433-022-1946-2

  21. [29]

    Luminet, J. P. 1979, A&A, 75, 228

  22. [30]

    R., Psaltis, D., & ¨Ozel, F

    Medeiros, L., Lauer, T. R., Psaltis, D., & ¨Ozel, F. 2018, ApJ, 864, 7, doi: 10.3847/1538-4357/aad37a

  23. [31]

    R., & ¨Ozel, F

    Medeiros, L., Psaltis, D., Lauer, T. R., & ¨Ozel, F. 2023, ApJL, 947, L7, doi: 10.3847/2041-8213/acc32d

  24. [32]

    M., Porth, O., et al

    Mizuno, Y., Younsi, Z., Fromm, C. M., Porth, O., et al. 2018, Nature Astronomy, 2, 585, doi: 10.1038/s41550-018-0449-5

  25. [33]

    2024, MNRAS, 527, 10965, doi: 10.1093/mnras/stad3797 Mo´ scibrodzka, M., & Gammie, C

    Mohan, A., Protopapas, P., Kunnumkai, K., Garraffo, C., et al. 2024, MNRAS, 527, 10965, doi: 10.1093/mnras/stad3797 Mo´ scibrodzka, M., & Gammie, C. F. 2018, MNRAS, 475, 43, doi: 10.1093/mnras/stx3162

  26. [34]

    A., & Yfantis, A

    Moscibrodzka, M. A., & Yfantis, A. I. 2023, ApJS, 265, 22, doi: 10.3847/1538-4365/acb6f9

  27. [35]

    2022, MNRAS, 511, 3795, doi: 10.1093/mnras/stac285

    Narayan, R., Chael, A., Chatterjee, K., Ricarte, A., et al. 2022, MNRAS, 511, 3795, doi: 10.1093/mnras/stac285

  28. [36]

    V., & Abramowicz, M

    Narayan, R., Igumenshchev, I. V., & Abramowicz, M. A. 2003, PASJ, 55, L69, doi: 10.1093/pasj/55.6.L69

  29. [37]

    1997, ApJ, 476, 49, doi: 10.1086/303591

    Narayan, R., Kato, S., & Honma, F. 1997, ApJ, 476, 49, doi: 10.1086/303591

  30. [38]

    2021, ApJ, 912, 35, doi: 10.3847/1538-4357/abf117

    Gelles, Z., et al. 2021, ApJ, 912, 35, doi: 10.3847/1538-4357/abf117

  31. [39]

    2021, PhRvD, 104, 124010, doi: 10.1103/PhysRevD.104.124010

    Peng, J., Guo, M., & Feng, X.-H. 2021, PhRvD, 104, 124010, doi: 10.1103/PhysRevD.104.124010

  32. [40]

    2016a, ApJ, 831, 4, doi: 10.3847/0004-637X/831/1/4

    Pu, H.-Y., Akiyama, K., & Asada, K. 2016a, ApJ, 831, 4, doi: 10.3847/0004-637X/831/1/4

  33. [41]

    Pu, H.-Y., & Broderick, A. E. 2018, ApJ, 863, 148, doi: 10.3847/1538-4357/aad086

  34. [42]

    2016b, ApJ, 820, 105, doi: 10.3847/0004-637X/820/2/105

    Pu, H.-Y., Yun, K., Younsi, Z., & Yoon, S.-J. 2016b, ApJ, 820, 105, doi: 10.3847/0004-637X/820/2/105

  35. [43]

    2022, ApJ, 938, 2, doi: 10.3847/1538-4357/ac8f49

    Qin, X., Chen, S., Zhang, Z., & Jing, J. 2022, ApJ, 938, 2, doi: 10.3847/1538-4357/ac8f49

  36. [44]

    N., et al

    Qiu, R., Ricarte, A., Narayan, R., Wong, G. N., et al. 2023, MNRAS, 520, 4867, doi: 10.1093/mnras/stad466

  37. [45]

    2024, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Ran, H., Guizilini, V., & Wang, Y. 2024, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

  38. [46]

    SaraerToosi, A., & Broderick, A. E. 2024, ApJ, 967, 140, doi: 10.3847/1538-4357/ad3e76

  39. [47]

    2024, IEEE Transactions on Medical Imaging, 43, 3559, doi: 10.1109/TMI.2024.3418652

    Shi, Y., Xia, W., Wang, G., & Mou, X. 2024, IEEE Transactions on Medical Imaging, 43, 3559, doi: 10.1109/TMI.2024.3418652

  40. [48]

    2014, Computer Science

    Simonyan, K., & Zisserman, A. 2014, Computer Science

  41. [49]

    D., & Kazemi, A

    Stephen, K. D., & Kazemi, A. 2014, Geophysical Prospecting, 62, 1009, doi: 10.1111/1365-2478.12109

  42. [50]

    2024, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition van der Gucht, J., Davelaar, J., Hendriks, L., Porth, O., et al

    Tu, S., Dai, Q., Cheng, Z.-Q., et al. 2024, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition van der Gucht, J., Davelaar, J., Hendriks, L., Porth, O., et al. 2020, A&A, 636, A94, doi: 10.1051/0004-6361/201937014

  43. [51]

    2011, Classical and Quantum Gravity, 28, 225011, doi: 10.1088/0264-9381/28/22/225011

    Perrin, G. 2011, Classical and Quantum Gravity, 28, 225011, doi: 10.1088/0264-9381/28/22/225011

  44. [52]

    2000, Journal of Statistical Planning and Inference, 86, 157

    Wan, A., & Ohtani, K. 2000, Journal of Statistical Planning and Inference, 86, 157

  45. [53]

    2022, Communications in Theoretical Physics, 74, 097401, doi: 10.1088/1572-9494/ac6e5c

    Wang, M., Chen, S., & Jing, J. 2022, Communications in Theoretical Physics, 74, 097401, doi: 10.1088/1572-9494/ac6e5c

  46. [54]

    2004, IEEE Transactions on Image Processing, 13, 600, doi: 10.1109/TIP.2003.819861

    Wang, Z., Bovik, A., Sheikh, H., & Simoncelli, E. 2004, IEEE Transactions on Image Processing, 13, 600, doi: 10.1109/TIP.2003.819861

  47. [55]

    Wang, Z., & Bovik, A. C. 2009, IEEE Signal Processing Magazine, 26, 98, doi: 10.1109/MSP.2008.930649 20

  48. [56]

    White, C. J. 2022, ApJS, 262, 28, doi: 10.3847/1538-4365/ac77ef

  49. [57]

    I., Zhao, S., Gold, R., Mo´ scibrodzka, M., et al

    Yfantis, A. I., Zhao, S., Gold, R., Mo´ scibrodzka, M., et al. 2024, MNRAS, 535, 3181, doi: 10.1093/mnras/stae2509

  50. [58]

    2014, ARA&A, 52, 529, doi: 10.1146/annurev-astro-082812-141003

    Yuan, F., & Narayan, R. 2014, ARA&A, 52, 529, doi: 10.1146/annurev-astro-082812-141003

  51. [59]

    2003, ApJ, 598, 301, doi: 10.1086/378716

    Yuan, F., Quataert, E., & Narayan, R. 2003, ApJ, 598, 301, doi: 10.1086/378716

  52. [60]

    2022, European Physical Journal C, 82, 835, doi: 10.1140/epjc/s10052-022-10794-z —

    Zhang, Z., Chen, S., & Jing, J. 2022, European Physical Journal C, 82, 835, doi: 10.1140/epjc/s10052-022-10794-z —. 2024, JCAP, 2024, 027, doi: 10.1088/1475-7516/2024/09/027

  53. [61]

    2024, JCAP, 05, 032, doi: 10.1088/1475-7516/2024/05/032

    Zhang, Z., Hou, Y., Guo, M., & Chen, B. 2024, JCAP, 05, 032, doi: 10.1088/1475-7516/2024/05/032

  54. [62]

    2025, in The Thirteenth International Conference on Learning Representations

    Zheng, Y., Liang, R., ZHENG, K., et al. 2025, in The Thirteenth International Conference on Learning Representations. https://openreview.net/forum?id=wM2sfVgMDH

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.