Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

CD-TVD: Contrastive Diffusion for 3D Super-Resolution with Scarce High-Resolution Time-Varying Data

T0 review · 5 major / 6 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 CD-TVD claims that a diffusion-based 3D super-resolution model, pre-trained on historical simulation data with contrastive degradation learning, can be adapted to a new simulation scenario using just one high-resolution timestep and then ac

desk verdict Plausible two-stage pre-train/fine-tune for 3D diffusion SR, but the 'one HR timestep' claim is undermined by an ambiguous evaluation protocol and lopsided baselines. read the letter →

arxiv 2508.08173 v2 pith:UGBQROZ6 submitted 2025-08-11 cs.CV eess.IV

Chongke Bi , Xin Gao , Jiangkang Deng , Guan Li , Jun Han This is my paper
classification cs.CVeess.IV
keywords 3Dsuper-resolutiontime-varyingdatadiffusionmodelcontrastivelearningscientificvisualizationfew-shotfine-tuningfluiddynamicslocalattention
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 tries to establish that super-resolving 3D time-varying simulation data does not require a large paired high/low-resolution dataset for every new scenario. Instead, a model pre-trained on historical simulations can learn general degradation patterns — how fine detail is lost when resolution drops — and then be fine-tuned for a new dataset with a single high-resolution timestep. The framework pairs a contrastive encoder, which pulls reconstructed volumes toward ground-truth high-resolution features and away from low-resolution ones, with a diffusion super-resolution network that recovers fine-grained structure. If true, this would make super-resolution practical for large-scale scientific simulations, where producing high-resolution data is often the most expensive step.

What carries the argument

The load-bearing mechanism is a two-module training loop. (1) A Contrastive Encoding Module maps 3D volumes into a latent space and is trained adversarially so that the restored volume $\varphi(I,w)$ is pulled toward the clear high-resolution volume $J$ and pushed away from the blurry low-resolution volume $I$, via the contrastive loss $\mathcal{L}_{\mathrm{CLD}}$. (2) An improved diffusion super-resolution network, built on a three-stage SwinUNet with Local Attention Blocks that apply window-based self-attention within non-overlapping 3D windows, is conditioned on the LR input and adds back high-frequency detail; it trains with 1000 diffusion steps and runs 20 steps at inference. An entropy

What would settle it

Concrete test: take a simulation dataset deliberately unlike the half-cylinder pre-training data (e.g., a different governing equation, very different resolution, or a different physical scale), fine-tune CD-TVD with exactly one entropy-selected high-resolution timestep, and compare reconstruction quality on held-out timesteps against PSRFlow trained with the same single timestep. If CD-TVD no longer clearly beats the baseline, or if its PSNR on target timesteps collapses while the baselines hold, the transfer premise is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the high-resolution-to-low-resolution degradation process is transferable across simulation scenarios, and can therefore be learned once and reused. CD-TVD pre-trains on historical simulation data (half-cylinder flow at three Reynolds numbers, roughly 450 timesteps) so the contrastive encoding module learns what distinguishes high-resolution from low-resolution fields, while the diffusion super-resolution network learns to add back missing high-frequency detail. For a new scenario, the encoder is frozen and only the diffusion module is fine-tuned on a single high-resolution timestep selected by maximum entropy, after which the model reconstructs all low-resolution t

Load-bearing premise

The load-bearing premise is that degradation patterns learned on historical simulation data transfer to a new scenario, so fine-tuning on one high-resolution timestep is enough; the paper's Discussion concedes that performance is sensitive to dataset similarity, so if the new dataset is too different from the pre-training distribution, one-timestep fine-tuning will not reconstruct accurately.

Editorial extensions

If this is right

  • Reduces the high-resolution data requirement for applying super-resolution to a new simulation from a large paired dataset to one timestep, cutting the costliest part of simulation workflows.
  • Makes diffusion-based super-resolution computationally feasible for 3D volumes: windowed local attention keeps memory and runtime low enough for a single GPU while retaining fine detail.
  • Amortizes the expensive pre-training (about 36 hours) across many new scenarios, each needing only a short fine-tuning step (about 5 hours).
  • Gives a principled rule for choosing which single timestep to simulate at high resolution: the timestep with the highest entropy in its low-resolution counterpart.
  • Extends the practical reach of super-resolution to large-scale fluid and atmospheric simulations where high-resolution ground truth is scarce.

Reading between the lines

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

  • A natural editorial extension: the same contrastive degradation prior should transfer to other inverse problems with scarce high-fidelity targets, such as temporal super-resolution, denoising, or sparse-sensor reconstruction, if the degradation operator is similar.
  • The paper's admitted sensitivity to dataset similarity suggests a testable design rule: measure the distribution gap between pre-training and target data before trusting one-timestep fine-tuning; a similarity threshold could be calibrated from the observed PSNR drop.
  • One could test whether entropy is actually the best selection criterion by comparing it with alternatives such as highest temporal gradient or largest local variance on the same four datasets; the paper does not compare selection strategies.
  • Because fine-tuning uses only one timestep, the method implicitly assumes a single snapshot carries enough information about the new scenario's degradation; for highly nonstationary simulations, a frame near a regime transition might be more informative than the maximum-entropy frame.
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

5 major / 6 minor

Summary. The paper proposes CD-TVD, a two-stage framework for 3D super-resolution of time-varying scientific simulation data under scarcity of high-resolution (HR) timesteps. In the pre-training stage, a contrastive encoder learns HR-LR degradation patterns on historical simulation data, while a diffusion-based super-resolution network with a local attention mechanism is trained to reconstruct fine details. In the fine-tuning stage, the contrastive module is frozen and only a single HR timestep is used to adapt the diffusion module to a new scenario, after selecting that timestep by maximum entropy of the LR fields. The method is evaluated on four datasets (Tangaroa, Half Cylinder Ensemble, Shock Interaction Vortex, Hurricane) against trilinear interpolation, SRGAN, SSR-VFD, and PSRFlow, using PSNR, LPIPS, and Chamfer distance. The main claim is that one HR timestep, combined with pre-trained degradation knowledge, suffices to reconstruct all other LR timesteps in a new dataset.

Significance. If the central claim holds, CD-TVD would be a practically useful contribution: it directly addresses the high cost of HR scientific simulation data and shows how pre-trained degradation knowledge can be repurposed with a single HR adaptation sample. The two-stage contrastive-plus-diffusion design is reasonable, and the inclusion of local attention to reduce 3D diffusion cost is sensible. The paper also makes its code available, uses multiple quantitative metrics, and includes ablation studies. However, the evidence as presented is not yet sufficient: the baseline comparisons are unequal in pre-training data, the evaluation protocol may include the fine-tuning timestep, the key-timestep selection heuristic is not validated, no variance or significance information is provided, and there are numerical inconsistencies between the main table and the ablation table. These issues directly affect the validity of the central claim, so the result is promising but needs substantial revision before it can be accepted.

major comments (5)
  1. [§4.1, §4.2.1, Table 3] The comparison is not controlled for pre-training data. CD-TVD is pre-trained on approximately 450 HR timesteps from the Half Cylinder ensemble (§4.1) and then fine-tuned with the single selected HR timestep, while SRGAN, SSR-VFD, and PSRFlow are trained from scratch on that same single timestep (§4.2.1). The margin reported in Table 3 could therefore be due to the much larger pre-training corpus rather than to the proposed single-timestep adaptation mechanism. To support the paper's central claim, please either pre-train the baselines on the same historical data, or train CD-TVD from scratch under the identical single-timestep protocol, and report both settings.
  2. [§3.3, §4.2.1, Table 3, Fig. 9] The evaluation may include the fine-tuning timestep. The entropy-selected timestep tmax from Eq. (4) is used as training data for CD-TVD and all baselines, and Table 3 reports metrics averaged 'over all timesteps.' The manuscript never states that tmax is excluded from the computed averages. If it is included, the reported PSNR/LPIPS/CD contain a training point, so the abstract and conclusion claim of reconstructing 'all subsequent LR timesteps' from one HR timestep is not supported by the numbers. Please explicitly exclude the fine-tuning timestep from evaluation, report held-out timestep metrics separately, and mark tmax in Fig. 9.
  3. [§3.3] The entropy-based key-timestep selection is an unverified assumption. The paper asserts that the highest-entropy LR timestep is the most informative for fine-tuning, but no experiment compares this choice against random selection, minimum entropy, or other selection rules, and there is no analysis of sensitivity to the selected timestep. Since the entire fine-tuning budget is a single HR timestep, the method's success is directly tied to this choice. Please add an ablation with random or fixed timesteps over multiple trials and report the distribution of held-out performance.
  4. [§4.1, §4.2, Table 3] No error bars or statistical significance tests are reported. Diffusion inference is stochastic, so a single run may not be representative. In addition, the hyperparameter β is grid-searched on the Tangaroa dataset (§4.1) and then Tangaroa is used for the main evaluation (Table 3) and ablation (Table 4), which is a form of tuning on the test set. Please report mean ± standard deviation over multiple seeds, and for β use a validation split or a separate dataset for selection.
  5. [Table 3 vs Table 4; §4.1] Two reporting issues undermine verifiability. First, Table 4 lists the full CD-TVD Tangaroa Chamfer distance as 1.3609, while Table 3 gives 1.9320 for the same configuration; one of these tables must be wrong, which casts doubt on the ablation conclusions. Second, CD-TVD is pre-trained on the Half Cylinder ensemble (§4.1) and the Half Cylinder dataset is included in the main evaluation (Table 3, Fig. 9). If the evaluation row is from the same simulation family, even at a different Reynolds number, it is not a 'new scenario' and should be labeled and analyzed separately. Please correct the numerical discrepancy and clarify the dataset relationship.
minor comments (6)
  1. [Abstract] Use 'we propose' rather than 'we proposed' in the abstract; also check similar tense issues elsewhere.
  2. [§3.3, Eq. (3)] The entropy formula is not well-defined for continuous volumetric data unless a binning or density-estimation procedure is specified. Please state how p(x_i) is estimated.
  3. [§4.1] The Half Cylinder description says the focus case has Reynolds number 6400, while the pre-training paragraph says the ensemble includes Reynolds numbers 160, 320, and 640. Clarify whether the tested Half Cylinder row is a separate Re=6400 case or one of the pre-training cases.
  4. [§4.2.1] The sentence 'we used five residual blocks (RB) and applied perceptual losses instead of the perceptual loss used in the original implementation' is self-contradictory. State exactly which loss was used for SRGAN.
  5. [§4.2.3, §4.3.3] The name 'SSR-TVD' appears in the cost comparison and discussion, but the baseline is called 'SSR-VFD' in the method list. Standardize the naming.
  6. [Fig. 9] The caption says 'synthesized vector fields' but the figure shows metric curves and box plots. Also label the rows explicitly rather than only in the caption.

Circularity Check

2 steps flagged · score 5.0 of 10

Reported all-timestep averages include the entropy-selected fine-tuning timestep, and beta is tuned on the Tangaroa test set; no derivation-level circularity, but the headline evidence is partly fitted.

  1. fitted input called prediction [Section 3.3, Section 4.2.1, Table 3]
    "Specifically, each baseline method (SRGAN, SSR-VFD, PSRFlow) was trained using the same single HR timestep, selected via our entropy-based selection method (detailed in Section 3.3). ... Table 3 presents the averaged PSNR, LPIPS, and CD scores over all timesteps for each dataset."

    By Eq. (4), tmax = argmax_t H(X_t) is the single timestep used for fine-tuning. The reported headline scores in Table 3 are 'averaged ... over all timesteps', and the paper never states that tmax is excluded. Hence the average contains a training point by construction: the model was fine-tuned to reconstruct this timestep, so its PSNR/LPIPS/CD in the average are fitted values, not predictions. The contribution claim 'reconstruct all subsequent LR timesteps' is therefore not established by an average that includes the fine-tuning timestep; a held-out evaluation separating the fine-tuning timestep from evaluated timesteps is required.

  2. fitted input called prediction [Section 4.1, Table 1, Table 3, Table 4]
    "To determine the optimal value of the hyper-parameter β in our loss function (Equation 1), we performed a grid search on the Tangaroa dataset. We systematically explored a range of values and evaluated the reconstruction performance based on Peak Signal-to-Noise Ratio (PSNR). ... Based on these results, we selected β=0.1 as the default setting, since it achieved the highest PSNR"

    The hyperparameter beta is selected by maximizing PSNR on the Tangaroa dataset, and the same dataset's CD-TVD row in Table 3 and the entire Table 4 ablation are then reported as experimental evidence. The Tangaroa numbers are therefore not independent predictions: they are partly selected by the beta grid search on that exact dataset. This is a localized evaluation-protocol circularity rather than a derivation-level one, because beta is a single scalar and the other datasets are not used for this selection.

full rationale

The paper is an empirical machine-learning framework; there is no mathematical derivation chain whose conclusions are equivalent to its inputs by definition. The many self-citations are related-work references and are not load-bearing for the central claim, and no uniqueness theorem or ansatz is imported from the authors' prior work. The two genuine issues are evaluation-protocol overlaps. First, the fine-tuning timestep selected via entropy (Section 3.3) is not explicitly excluded from the 'averaged over all timesteps' metrics in Table 3, so the headline averages may contain a training point by construction. Second, the beta hyperparameter is grid-searched on the Tangaroa dataset and then Tangaroa is used again for final evaluation and ablation, making those specific numbers partially fitted rather than predicted. Both issues are localized: predictions on other timesteps and other datasets remain genuine empirical evaluations. The paper's own Discussion admits sensitivity to dataset similarity, which weakens the universality claim but is a limitation, not circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim relies on the transferability of learned degradation patterns across datasets, which presupposes a consistent downsampling model and a representative fine-tuning timestep. These are not external benchmarks but choices made by the authors, and they directly affect the measured performance.

free parameters (2)
  • beta (β) = 0.1
    Hyperparameter balancing reconstruction and contrastive loss, chosen by grid search on Tangaroa dataset (Table 1) and then applied to all datasets.
  • Inference diffusion steps = 20
    Number of reverse diffusion steps at inference, set to 20 based on empirical balance between quality and efficiency (Section 3.2).
assumptions (3)
  • domain assumption LR data is generated from HR via trilinear downsampling
    Used to create all training and evaluation pairs; real-world LR simulations may have different degradation processes. Section 4.1.
  • domain assumption The contrastive loss (Eq. 2) encourages the SR output to be closer to HR than LR in the learned embedding space
    The paper does not prove that this loss is well-calibrated for the stated goal; the notation is ambiguous, and the roles of IHR, ILR, ISR in the two expectation terms are inconsistent. Section 3.1.1.
  • ad hoc to paper The highest-entropy LR timestep is the most informative for fine-tuning
    Only intuitive justification is given; no ablation compares entropy selection with random or other criteria. Section 3.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CD-TVD: Contrastive Diffusion for 3D Super-Resolution with Scarce High-Resolution Time-Varying Data." pith.science (2026). https://pith.science/paper/UGBQROZ6

@misc{pith2026250808173,
  author       = {Pith},
  title        = {Pith review of: CD-TVD: Contrastive Diffusion for 3D Super-Resolution with Scarce High-Resolution Time-Varying Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UGBQROZ6}},
  note         = {Machine review of arXiv:2508.08173}
}
read the original abstract

Large-scale scientific simulations require significant resources to generate high-resolution time-varying data (TVD). While super-resolution is an efficient post-processing strategy to reduce costs, existing methods rely on a large amount of HR training data, limiting their applicability to diverse simulation scenarios. To address this constraint, we proposed CD-TVD, a novel framework that combines contrastive learning and an improved diffusion-based super-resolution model to achieve accurate 3D super-resolution from limited time-step high-resolution data. During pre-training on historical simulation data, the contrastive encoder and diffusion superresolution modules learn degradation patterns and detailed features of high-resolution and low-resolution samples. In the training phase, the improved diffusion model with a local attention mechanism is fine-tuned using only one newly generated high-resolution timestep, leveraging the degradation knowledge learned by the encoder. This design minimizes the reliance on large-scale high-resolution datasets while maintaining the capability to recover fine-grained details. Experimental results on fluid and atmospheric simulation datasets confirm that CD-TVD delivers accurate and resource-efficient 3D super-resolution, marking a significant advancement in data augmentation for large-scale scientific simulations. The code is available at https://github.com/Xin-Gao-private/CD-TVD.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. TeDiO: Temporal Diagonal Optimization for Training-Free Coherent Video Diffusion

    cs.CV 2026-05 unverdicted novelty 7.0 of 10

    TeDiO regularizes temporal diagonals in diffusion transformer attention maps to produce smoother video motion while keeping per-frame quality intact.

Reference graph

Works this paper leans on

60 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    An, H.-W

    Y . An, H.-W. Shen, G. Shan, G. Li, and J. Liu. Stsrnet: Deep joint space-time super-resolution for vector field visualization.IEEE Computer Graphics and Applications, 41(6):122–132, 2021. doi: 10.1109/MCG. 2021.3097555 2

    arXiv 2021

  2. [2]

    H. G. Barrow, J. M. Tenenbaum, R. C. Bolles, and H. C. Wolf. Para- metric correspondence and chamfer matching: Two new techniques for image matching. InProceedings of the International Joint Conference on Artificial Intelligence, p. 659–663, 1977. doi: doi/abs/10.5555/1622943. 1622971 6

    work page doi:10.5555/1622943 1977

  3. [3]

    H. Cao, C. Tan, Z. Gao, Y . Xu, G. Chen, P.-A. Heng, and S. Z. Li. A survey on generative diffusion models.IEEE Transactions on Knowledge and Data Engineering, 36(7):2814–2830, 2024. doi: 10.1109/TKDE.2024 .3361474 2

    doi:10.1109/tkde.2024 2024

  4. [4]

    G. Chen, B. Dong, Y . Zhang, W. Lin, D. Shen, and P.-T. Yap. Xq-sr: Joint x-q space super-resolution with application to infant diffusion mri. Medical Image Analysis, 57:44–55, 2019. doi: 10.1016/j.media.2019.06. 010 2

    work page doi:10.1016/j.media.2019.06 2019

  5. [5]

    Chung, E

    H. Chung, E. S. Lee, and J. C. Ye. Mr image denoising and super-resolution using regularized reverse diffusion.IEEE Transactions on Medical Imag- ing, 42(4):922–934, 2023. doi: 10.1109/TMI.2022.3220681 2

    work page arXiv 2023

  6. [6]

    Croitoru, V

    F.-A. Croitoru, V . Hondru, R. T. Ionescu, and M. Shah. Diffusion models in vision: A survey.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 45(9):10850–10869, 2023. doi: 10.1109/TPAMI.2023. 3261988 2

    doi:10.1109/tpami.2023 2023

  7. [7]

    Daniels, T

    M. Daniels, T. Maunu, and P. Hand. Score-based generative neural net- works for large-scale optimal transport. InProceedings of Advances in Neural Information Processing Systems, pp. 12955–12965, 2021. doi: doi/abs/10.5555/3540261.3541253 2

    work page arXiv 2021

  8. [8]

    Z. Deng, C. He, Y . Liu, and K. C. Kim. Super-resolution reconstruction of turbulent velocity fields using a generative adversarial network-based artificial intelligence framework.Physics of Fluids, 31(12):125111, 2019. doi: 10.1063/1.5127031 2

    doi:10.1063/1.5127031 2019

Show all 60 references
  1. [9]

    Dhariwal and A

    P. Dhariwal and A. Nichol. Diffusion models beat gans on image synthesis. InProceedings of Advances in Neural Information Processing Systems, pp. 8780–8794, 2021. doi: 10.48550/arXiv.2105.05233 2

    2021 doi arXiv

  2. [10]

    H. Gao, X. Han, X. Fan, L. Sun, L.-P. Liu, L. Duan, and J.-X. Wang. Bayesian conditional diffusion models for versatile spatiotemporal turbu- lence generation.Computer Methods in Applied Mechanics and Engineer- ing, 427:117023, 2024. doi: 10.1016/j.cma.2024.117023 2

    2024

  3. [11]

    S. Gao, X. Liu, B. Zeng, S. Xu, Y . Li, X. Luo, J. Liu, X. Zhen, and B. Zhang. Implicit diffusion models for continuous super-resolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 10021–10030, 2023. doi: 10.48550/arXiv.2303.16491 2

    2023 doi arXiv

  4. [12]

    L. Guo, S. Ye, J. Han, H. Zheng, H. Gao, D. Z. Chen, J.-X. Wang, and C. Wang. Ssr-vfd: Spatial super-resolution for vector field data analysis and visualization. InProceedings of IEEE Pacific Visualization Symposium, pp. 71–80, 2020. doi: 10.1109/PacificVis48177.2020.8737 2, 6

    2020

  5. [13]

    Han and C

    J. Han and C. Wang. Tsr-tvd: Temporal super-resolution for time-varying data analysis and visualization.IEEE Transactions on Visualization and Computer Graphics, 26(1):205–215, 2019. doi: 10.1109/TVCG.2019. 2934255 2

    2019 doi

  6. [14]

    Han and C

    J. Han and C. Wang. Ssr-tvd: Spatial super-resolution for time-varying data analysis and visualization.IEEE Transactions on Visualization and Computer Graphics, 28(6):2445–2456, 2020. doi: 10.1109/TVCG.2020. 3032123 2

    2020 doi

  7. [15]

    Han and C

    J. Han and C. Wang. Tsr-vfd: Generating temporal super-resolution for unsteady vector field data.Computers & Graphics, 103(1):168–179, 2022. doi: 10.1016/j.cag.2022.02.001 2

    2022 doi

  8. [16]

    J. Han, H. Zheng, D. Z. Chen, and C. Wang. STNet: An end-to-end generative framework for synthesizing spatiotemporal super-resolution volumes.IEEE Transactions on Visualization and Computer Graphics, 28(1):270–280, 2021. doi: 10.1109/TVCG.2021.3114815 2

    2021

  9. [17]

    J. Ho, C. Saharia, W. Chan, D. J. Fleet, M. Norouzi, and T. Salimans. Cascaded diffusion models for high fidelity image generation.Journal of Machine Learning Research, 23(47):1–33, 2022. doi: 10.48550/arXiv. 2106.15282 2

    2022 doi arXiv

  10. [18]

    J. Ho, T. Salimans, A. Gritsenko, W. Chan, M. Norouzi, and D. J. Fleet. Video diffusion models.In Proceedings of Advances in Neural Information Processing Systems, 35(6):8633–8646, 2022. doi: 10.48550/arXiv.2204. 03458 2

    2022 doi

  11. [19]

    C. Jiao, C. Bi, and L. Yang. FFEINR: Flow feature-enhanced implicit neural representation for spatiotemporal super-resolution.Journal of Visualization, 27(2):273–289, 2024. doi: 10.1007/s12650-024-00959-1 2

    2024 doi

  12. [20]

    C. Jiao, C. Bi, L. Yang, Z. Wang, Z. Xia, and K. Ono. Esrgan-based visu- alization for large-scale volume data.Journal of Visualization, 26(3):649– 665, 2023. doi: 10.1007/s12650-022-00891-2 2

    2023 doi

  13. [21]

    Karatsiolis, C

    S. Karatsiolis, C. Padubidri, and A. Kamilaris. Exploiting digital surface models for inferring super-resolution for remotely sensed images.IEEE Transactions on Geoscience and Remote Sensing, 60(1):1–13, 2022. doi: 10.1109/TGRS.2022.3209340 2

    2022

  14. [22]

    G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang. Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021. doi: 10.1038/s42254-021-00314-5 2

    2021 doi

  15. [23]

    D. P. Kingma. Adam: A method for stochastic optimization. InProceed- ings of International Conference on Learning Representations, 2015. doi: 10.48550/arXiv.1412.6980 5

    2015 doi arXiv

  16. [24]

    Ledig, L

    C. Ledig, L. Theis, F. Huszár, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, and W. Shi. Photo-realistic single image super-resolution using a generative adversarial network. InProceed- ings of IEEE Conference on Computer Vision and Pattern Re...

    2017 doi

  17. [25]

    D. C. Lepcha, B. Goyal, A. Dogra, and V . Goyal. Image super-resolution: A comprehensive review, recent trends, challenges and applications.Infor- mation Fusion, 91(1):230–260, 2023. doi: 10.1016/j.inffus.2022.10.007 2

    2023 doi

  18. [26]

    H. Li, Y . Yang, M. Chang, S. Chen, H. Feng, Z. Xu, Q. Li, and Y . Chen. SRDiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479(1):47–59, 2022. doi: 10.1016/j.neucom.2022.01. 029 2

    2022 doi

  19. [27]

    J. Liu, Z. Yuan, Z. Pan, Y . Fu, L. Liu, and B. Lu. Diffusion model with detail complement for super-resolution of remote sensing.Remote Sensing, 14(19):4834, 2022. doi: 10.3390/rs14194834 2

    2022 doi

  20. [28]

    K. Liu, C. Jiao, X. Gao, and C. Bi. Uginr: Large-scale unstructured grid reduction via implicit neural representation.Journal of Visualization, 27(5):983–996, 2024. doi: 10.1007/s12650-024-01003-y 2

    2024 doi

  21. [29]

    P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, and G. Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing.ACM computing surveys, 55(9):1–35, 2023. doi: 10. 1145/3560815 2

    2023

  22. [30]

    Metzger, R

    N. Metzger, R. C. Daudt, and K. Schindler. Guided depth super-resolution by deep anisotropic diffusion. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 18237–18246, 2023. doi: 10.48550/arXiv.2211.11592 2

    2023 doi arXiv

  23. [31]

    A. Q. Nichol and P. Dhariwal. Improved denoising diffusion probabilistic models. InProceedings of International Conference on Machine Learning, pp. 8162–8171, 2021. doi: 10.48550/arXiv.2102.09672 2

    2021 doi arXiv

  24. [32]

    L. Ning, K. Setsompop, O. Michailovich, N. Makris, M. E. Shenton, C.-F. Westin, and Y . Rathi. A joint compressed-sensing and super- resolution approach for very high-resolution diffusion imaging.Neu- roImage, 125(1):386–400, 2016. doi: 10.1016/j.neuroimage.2015.10.061 2

    2016 doi

  25. [33]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 10684–10695, 2022. doi: 10.48550/arXiv.2112.10752 2

    2022 doi arXiv

  26. [34]

    Saharia, J

    C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi. Image super-resolution via iterative refinement.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713–4726, 2022. doi: 10.1109/TPAMI.2022.3204461 2

    2022

  27. [35]

    Schanz, F

    A. Schanz, F. List, and O. Hahn. Stochastic super-resolution of cosmolog- ical simulations with denoising diffusion models.The Open Journal of Astrophysics, 7(8), 2024. doi: 10.33232/001c.125902 2

    2024 doi

  28. [36]

    Shen and H.-W

    J. Shen and H.-W. Shen. PSRFlow: Probabilistic super resolution with flow-based models for scientific data.IEEE Transactions on Visualization and Computer Graphics, 30(3):986–996, 2023. doi: 10.1109/TVCG.2023. 3327171 2, 6

    2023 doi

  29. [37]

    L. Shen, L. Deng, X. Liu, Y . Wang, X. Chen, and J. Liu. A generative adversarial network based on an efficient transformer for high-fidelity flow field reconstruction.Physics of Fluids, 36(7), 2024. doi: 10.1063/5. 0215681 1

    2024 doi

  30. [38]

    L. Shen, L. Deng, Y . Wang, J. Zhang, and J. Liu. Pcsagan: A physics- constrained generative network based on self-attention for high-fidelity flow field reconstruction.Journal of Visualization, 27(4):661–676, 2024. doi: 10.1007/s12650-024-00987-x 1

    2024 doi

  31. [39]

    J. Song, Z. Song, P. Ren, N. B. Erichson, M. W. Mahoney, and X. S. Li. Forecasting high-dimensional spatio-temporal systems from sparse measurements.Machine Learning: Science and Technology, 5(4):045067,

  32. [40]

    X. Song, G. Wang, W. Zhong, K. Guo, Z. Li, X. Liu, J. Dong, and Q. Liu. Sparse-view reconstruction for photoacoustic tomography combining dif- fusion model with model-based iteration.Photoacoustics, 33(1):100558,

  33. [41]

    G. Vis, M. Nilsson, C.-F. Westin, and F. Szczepankiewicz. Accuracy and precision in super-resolution mri: Enabling spherical tensor diffu- sion encoding at ultra-high b-values and high resolution.NeuroImage, 245(2):118673, 2021. doi: 10.1016/j.neuroimage.2021.118673 2

    2021

  34. [42]

    Wang and J

    C. Wang and J. Han. Dl4scivis: A state-of-the-art survey on deep learn- ing for scientific visualization.IEEE Transactions on Visualization and Computer Graphics, 29(8):3714–3733, 2022. doi: 10.1109/TVCG.2022. 3167896 1

    2022 doi

  35. [43]

    M. Wang, C. Bi, L. Yang, X. Qiu, Y . Li, and C. Yu. Pmim: Generating high-resolution air pollution data via masked image modeling.Journal of Visualization, 27(3):383–399, 2024. doi: 10.1007/s12650-024-00965-3 2

    2024 doi

  36. [44]

    P. Wang, B. Bayram, and E. Sertel. A comprehensive review on deep learning based remote sensing image super-resolution methods.Earth- Science Reviews, 232(1):104110, 2022. doi: 10.1016/j.earscirev.2022. 104110 2

    2022 doi

  37. [45]

    X. Wang, Y . Dong, S. Zou, L. Zhang, and X. Deng. A semi-supervised framework for computational fluid dynamics prediction.Applied Soft Computing, 154(6):111422, 2024. doi: 10.1016/j.asoc.2024.111422 2

    2024

  38. [46]

    Z. Wang, J. Chen, and S. C. Hoi. Deep learning for image super-resolution: A survey.IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 43(10):3365–3387, 2020. doi: 10.1109/TPAMI.2020.2982166 2

    2020

  39. [47]

    G. Wu, J. Jiang, and X. Liu. A practical contrastive learning framework for single-image super-resolution.IEEE Transactions on Neural Networks and Learning Systems, 35(3):15834–15845, 2024. doi: 10.1109/TNNLS. 2023.3290038 4

    2024

  40. [48]

    Z. Wu, X. Chen, S. Xie, J. Shen, and Y . Zeng. Super-resolution of brain mri images based on denoising diffusion probabilistic model.Biomedical Signal Processing and Control, 85(1):104901, 2023. doi: 10.1016/j.bspc. 2023.104901 2

    2023

  41. [49]

    S. W. Wurster, H. Guo, H.-W. Shen, T. Peterka, and J. Xu. Deep hierarchi- cal super resolution for scientific data.IEEE Transactions on Visualization and Computer Graphics, 29(12):5483–5495, 2023. doi: 10.1109/TVCG. 2022.3214420 2

    2023

  42. [50]

    Z. Xiao, K. Kreis, and A. Vahdat. Tackling the generative learning trilemma with denoising diffusion GANs. InProceedings of International Conference on Learning Representations, 2022. doi: 10.48550/arXiv.2112 .07804 2

    2022 doi

  43. [51]

    Y . Xie, E. Franz, M. Chu, and N. Thuerey. TempoGAN: A temporally coherent, volumetric GAN for super-resolution fluid flow.ACM Trans- actions on Graphics, 37(4):1–15, 2018. doi: 10.1145/3197517.3201304 2

    2018

  44. [52]

    L. Yang, Z. Zhang, S. Hong, R. Xu, Y . Zhao, Y . Shao, W. Zhang, M.-H. Yang, and B. Cui. Diffusion models: A comprehensive survey of methods and applications.ACM Computing Surveys, 56(4):1–39, 2022. doi: 10. 1145/3626235 2

    2022

  45. [53]

    W. Yang, X. Zhang, Y . Tian, W. Wang, J.-H. Xue, and Q. Liao. Deep learning for single image super-resolution: A brief review.IEEE Transac- tions on Multimedia, 21(12):3106–3121, 2019. doi: 10.1109/TMM.2019. 2919431 2

    2019 doi

  46. [54]

    Y . Yang, C. Jiao, X. Gao, X. Tian, and C. Bi. Adaptive volumetric data compression based on implicit neural representation. InProceedings of the International Symposium on Visual Information Communication and Interaction, pp. 1–8, 2024. doi: 10.1145/3678698.3678703 1

    2024

  47. [55]

    Z. Yue, J. Wang, and C. C. Loy. Resshift: Efficient diffusion model for image super-resolution by residual shifting. InProceedings of Advances in Neural Information Processing Systems, pp. 13294–13307, 2023. doi: 10.48550/arXiv.2307.12348 2

    2023 doi arXiv

  48. [56]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreason- able effectiveness of deep features as a perceptual metric. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 586–595, 2018. doi: 10.1109/CVPR.2018.00068 6

    2018

  49. [57]

    Z. Zhou, Y . Hou, Q. Wang, G. Chen, J. Lu, Y . Tao, and H. Lin. V olume upscaling with convolutional neural networks. InProceedings of the Computer Graphics International Conference, pp. 1–6, 2017. doi: 10. 1145/3095140.3095178 2

    2017

  50. [58]

    Z. Zuo, T. Fang, H. Wu, and Z. Zhang. High-resolution reconstruction algorithm for the three-dimensional velocity field produced by atomiza- tion of two impinging jets based on deep learning.Physics of Fluids, 35(6):063306, 2023. doi: 10.1063/5.0152779 2, 3

    2023 doi

  51. [2023]

    doi: 10.1016/j.pacs.2023.100558 2

    2023

  52. [2024]

    doi: 10.1088/2632-2153/ad9883 2

    doi

Pith tools

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