Pith. sign in

REVIEW 3 major objections 6 minor 44 references

Generative Latent Diffusion for Efficient Spatiotemporal Data Reduction

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A latent diffusion codec that stores only keyframes and generates the rest compresses scientific spatiotemporal data 4–10× better than rule-based compressors at equal reconstruction error.

desk verdict A genuinely new keyframe-only latent diffusion compression idea with consistent experimental curves, undermined only by an underspecified PCA post-processing step. read the letter →

arxiv 2507.02129 v1 pith:HOEOMCMX submitted 2025-07-02 cs.LG cs.CV

Xiao Li , Liangji Zhu , Anand Rangarajan , Sanjay Ranka This is my paper
classification cs.LGcs.CV
keywords latentdiffusionscientificdatacompressionspatiotemporalreductionkeyframeinterpolationerror-boundedvariationalautoencoderPCApost-processingconditional
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

Scientific simulations produce enormous spatiotemporal datasets, and this paper argues that a generative model can shrink them by storing only a few keyframes. A variational autoencoder compresses those keyframes into compact latent codes, and a conditional diffusion model generates the intermediate frames by denoising random noise while looking at the keyframe codes. A low-rank residual correction step then removes the remaining reconstruction error so the final output satisfies a user-specified error bound. On climate, combustion, and turbulence benchmarks, the method reports 4–10× higher compression than rule-based compressors such as SZ3 and 20–63% better than learning-based baselines at the same reconstruction error.

What carries the argument

The load-bearing mechanism is keyframe-conditioned latent diffusion: a denoising U-Net with factorized space-time attention operates on 64-channel latent tensors, receives the entire frame sequence with noise added only to frames that must be generated, and is trained with the loss restricted to those generated frames. The other half of the machinery is the PCA residual post-processor, which takes the decompressed block, computes the residual against the original, projects it onto a basis $\mathbf{U}$, and stores only the top $M$ quantized coefficients needed to push the $\ell^2$ error below the threshold $\tau$.

What would settle it

Recompute the compression ratio with the PCA basis and quantized residual coefficients fully included in the stored side information, and check whether the corrected ratios on E3SM, S3D, and JHTDB still beat SZ3 and ZFP at equal NRMSE. An additional check is to evaluate on time segments that were not used to train the VAE and diffusion model, ruling out in-sample fitting of the correction basis.

Watch

Extended reading notes

Core claim

The paper's central claim is that a conditional latent diffusion model can act as a learned spatiotemporal interpolator accurate enough for scientific compression: once keyframe latents are stored, every non-keyframe latent can be generated instead of stored, and the small residual after decoding is corrected by a PCA basis so the final reconstruction meets a guaranteed error bound. Because the diffusion model is trained with many denoising steps and then fine-tuned with as few as 32 steps, the generated frames are both storage-efficient and computationally practical. The authors claim this yields up to 10× better compression than rule-based compressors and 20–63% better compression than learning-based baselines on the E3SM, S3D, and JHTDB datasets.

Load-bearing premise

The reported compression gains depend on the low-rank correction basis being cheap to obtain and store; the paper does not state where that basis comes from, and if it is computed from the very data being compressed and stored alongside it, that storage is missing from the reported ratios.

Editorial extensions

If this is right

  • Storing only keyframe latents, rather than a latent for every frame, yields 4–10× higher compression than rule-based compressors at equal NRMSE on climate, combustion, and turbulence data.
  • Compared with leading learning-based codecs, the method improves compression ratio by 20–63% at the same reconstruction error.
  • Fine-tuning the diffusion model from 1,000 to 32 denoising steps preserves reconstruction accuracy while making decoding far faster, making the approach practical for large-scale data.
  • The PCA post-processing step converts a generative reconstruction into an error-bounded one, allowing the method to serve scientific workflows that require quantifiable accuracy on primary data.
  • Interpolation-based keyframe selection outperforms prediction-based and mixed strategies, with the optimal keyframe interval depending on the temporal correlation of the dataset.

Reading between the lines

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

  • The paper leaves implicit that its storage savings come from replacing per-frame latent storage with a trained conditional generator, so the same argument should transfer to other generative backbones whose decoders can be made fast; the diffusion model itself is not the only source of the ratio gain.
  • The optimal keyframe interval of 3 is a domain-specific hyperparameter, so a natural extension is to pick keyframes adaptively from temporal autocorrelation or motion estimates instead of uniform sampling.
  • If the low-rank correction basis is fixed and pretrained, the error-bound guarantee is nearly free at decode time; if it is adaptive, future comparisons should report its side-information cost explicitly, since that cost determines whether the reported 4–10× gaps survive.
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

3 major / 6 minor

Summary. The paper proposes a generative compression framework for spatiotemporal scientific data. A VAE with a hyperprior compresses selected keyframes into quantized latent codes; a conditional latent diffusion model, conditioned on those keyframe latents, generates latent representations for the remaining frames; and a PCA-based post-processing step is applied to the decoded residuals in an attempt to enforce an error bound. The authors report compression-ratio versus NRMSE curves on E3SM, S3D, and JHTDB data, claiming 4–10x improvement over SZ3/ZFP and 20–63% improvement over the learning-based baselines CDC, GCD, and VAE-SR, together with decoding-speed advantages from operating in latent space and from fine-tuning the diffusion model to few denoising steps.

Significance. If the claims hold, the paper would make a useful contribution: it couples a learned generative interpolator with transform coding in a way that stores only keyframe latents, and it provides a concrete speed argument for latent-space diffusion in scientific data reduction. The ablations on keyframe placement and denoising steps are informative, and the comparison across three distinct scientific datasets is a strength. However, the central quantitative claims currently depend on an incompletely specified PCA side-information term and on a non-rigorous error-bound argument, so the reported compression ratios are not yet interpretable as stated.

major comments (3)
  1. [Section 3.5, Eq. (10)] The stated error-bound guarantee is not established. The text says coefficients are selected so that the ℓ2 norm of the residual falls below a threshold τ, but the final reconstruction uses quantized coefficients c_q, not the unquantized coefficients c. Quantization changes the residual, so the pre-quantization selection does not imply ||x - x_G||_2 ≤ τ for the final reconstruction. A rigorous guarantee requires either selecting coefficients after accounting for quantization error, verifying the bound on the quantized reconstruction, or explicitly stating that the bound is empirical rather than guaranteed. As written, the 'guarantee' language in Sections 1, 3.5, and 5 is unsupported.
  2. [Section 3.5 and Section 4.1, Eq. (11)] The provenance and storage cost of the PCA basis U_s are unspecified, and this directly affects the headline compression ratios. The decoder must possess U_s to invert x_G = x_R + U_s c_q. If U_s is computed adaptively from the residual x - x_R for each block or dataset, its size must be included in Size(G) in Eq. (11); otherwise the decoder cannot reconstruct the data. If U_s is a fixed pretrained basis shared by encoder and decoder, that fact and the training procedure must be stated. If U_s is computed from test residuals, the evaluation is in-sample. The paper never reports what G contains or how large U_s is, so the 4–10x and 20–63% improvement figures are not yet verifiable.
  3. [Section 4.2 and Fig. 3] The experimental evaluation reports NRMSE, an average error metric, and does not report the achieved maximum ℓ2 residual against the target threshold τ. Since the paper's scientific-use claim rests on error-bound guarantees, the paper should report the actual worst-case or maximum residual for the post-processed reconstructions and compare it with τ. Without this, the claim that the method is suitable for error-bounded scientific workflows is not empirically demonstrated.
minor comments (6)
  1. [Section 2] There is a typo in 'evey block'; it should be 'every block'.
  2. [Section 3.4] The sentence 'the dimensionality of 𝒚 is significantly smaller than that of 𝒛' appears backwards; in a hyperprior model, the hyper-latent 𝒛 is typically lower-dimensional than 𝒚. Please clarify.
  3. [Table 2] The table heading says 'RTX 2080 24GB' but the text refers to an 'RTX 2090 24GB'; the GPU model name should be made consistent.
  4. [Figure 3] The legend labels for the CDC variants are truncated ('CDC-' rather than 'CDC-X' and 'CDC-ε'); the figure should be regenerated with complete labels.
  5. [Section 4.8] The exclusion of error-bound post-processing overhead from the speed comparison is a limitation that should be stated earlier and, ideally, accompanied by at least one end-to-end measurement; as written, the speedups are model-inference-only.
  6. [Section 3.5] The procedure for selecting the top M coefficients is described only qualitatively ('based on their contribution to the error'); please provide the actual selection algorithm and the bit-allocation rule used in the experiments.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: compression ratios are empirical comparisons against external baselines; the PCA post-processing is a standard transform-coding step with an under-specified side-information detail, not a circular reduction.

full rationale

The paper's central claim (up to 10x over SZ3, 20-63% over learned baselines) is supported by measured rate-distortion curves on E3SM, S3D, and JHTDB, comparing against external methods (SZ3, ZFP, CDC, GCD, VAE-SR). No equation defines a target in terms of itself, and no fitted parameter is relabeled as a prediction. The PCA error-bound post-processing (Eqs. 9-10) computes coefficients from the residual and reconstructs x_G = x_R + U_s c_q; this is a legitimate transform/post-processing construction, with the bit cost of coefficients intended to be counted in Size(G) in Eq. (11). The paper does not explicitly state whether the PCA basis U_s is a fixed pretrained/shared basis or is computed adaptively and transmitted; if U_s is omitted from Size(G), the reported ratios are not established. This is a side-information accounting/completeness concern, not a circularity, because the text does not define the ratio in terms of the reconstruction error or fit the error itself as the predicted quantity. Self-citations ([19,21,22,25] for PCA; [20] as GCD baseline) are not load-bearing: the PCA mechanism is described in the paper, and the numerical comparisons are external and measurable. Therefore the derivation chain is self-contained and non-circular.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities. The free parameters are training and inference hyperparameters chosen by hand, plus the unspecified PCA coefficient selection. The key axioms are the low-rank residual assumption and the availability of the PCA basis at decode time; both are unstated and load-bearing for the compression-ratio claim.

free parameters (5)
  • Rate-distortion Lagrange multiplier lambda = 1e-5, doubled at 250K iterations
    Balances MSE and bit-rate in the VAE loss (Eq. 8); chosen without a reported search, but it directly sets the operating point of the compression curve.
  • Keyframe interval (K) = 3 (optimal in ablation)
    Selected on E3SM as the best trade-off and applied to all datasets; the paper notes it is domain-specific, so it is a hand-picked hyperparameter.
  • Number of denoising steps at inference = 32
    Fine-tuned from 1000 steps; ablation shows 32 steps match 1000-step performance while being faster.
  • PCA coefficient selection threshold or rank M = not reported
    The post-processing selects top M PCA coefficients so that the residual norm is below tau; M varies with data and directly determines the side-information size, yet the paper gives no details on how M is chosen or priced.
  • Latent min-max normalization statistics = not specified
    Algorithm 1 normalizes latents to [-1,1]; the storage or transmission of these statistics is not addressed, and they are needed at inference.
assumptions (5)
  • domain assumption The residual after reconstruction is low-rank enough that a few PCA components can reduce the error below the target bound.
    Section 3.5 relies on this for the error-bound guarantee, but no analysis of residual rank is given.
  • domain assumption The PCA basis U is available to the decoder without incurring storage cost comparable to the compression savings.
    Eq. (10) requires U at decode time; the paper does not specify whether U is pretrained and fixed or adaptive and stored.
  • domain assumption The conditional diffusion model, trained on quantized and min-max-normalized latents, generalizes to unseen data blocks from the same domain.
    The compression quality depends on this generalization; no cross-domain or distribution-shift experiments are provided.
  • standard math Standard VAE and diffusion training objectives are used as in prior work and optimize the intended rate-distortion trade-off.
    Eqs. (7) and (8) follow from [30] and [15], and the paper provides no new theoretical justification.
  • standard math The entropy model (factorized density for z, Gaussian with uniform noise for y) correctly estimates the bit-rate of the quantized latents.
    Section 3.1 adopts the hyperprior model of [30]; this is prior art, not proven in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative Latent Diffusion for Efficient Spatiotemporal Data Reduction." pith.science (2026). https://pith.science/paper/HOEOMCMX

@misc{pith2026250702129,
  author       = {Pith},
  title        = {Pith review of: Generative Latent Diffusion for Efficient Spatiotemporal Data Reduction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HOEOMCMX}},
  note         = {Machine review of arXiv:2507.02129}
}
read the original abstract

Generative models have demonstrated strong performance in conditional settings and can be viewed as a form of data compression, where the condition serves as a compact representation. However, their limited controllability and reconstruction accuracy restrict their practical application to data compression. In this work, we propose an efficient latent diffusion framework that bridges this gap by combining a variational autoencoder with a conditional diffusion model. Our method compresses only a small number of keyframes into latent space and uses them as conditioning inputs to reconstruct the remaining frames via generative interpolation, eliminating the need to store latent representations for every frame. This approach enables accurate spatiotemporal reconstruction while significantly reducing storage costs. Experimental results across multiple datasets show that our method achieves up to 10 times higher compression ratios than rule-based state-of-the-art compressors such as SZ3, and up to 63 percent better performance than leading learning-based methods under the same reconstruction error.

Figures

Figures reproduced from arXiv: 2507.02129 by the authors.

Figure 1
Figure 1. Overview of the proposed architecture. Given an input spatiotemporal data block [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Comparison of keyframe selection strategies: interpolation, prediction, and mixed. The top images display the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Comparison of results on the E3SM, S3D and JHTDB datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Ablation Study for Denoising Step on S3D Dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 4
Figure 4. Figure 4: Ablation Study for interpolation interval on E3SM [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Visualization of reconstructed data using our method, VAE-SR, CDC, SZ3, and ZFP compressors. All methods compress [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 27 canonical work pages

  1. [1]

    Mark Ainsworth, Ozan Tugluk, Ben Whitney, and Scott Klasky. 2018. Multilevel techniques for compression and reduction of scientific data - the univariate case. Computing and Visualization in Science 19, 5-6 (2018), 65–76

    work page 2018

  2. [2]

    Mark Ainsworth, Ozan Tugluk, Ben Whitney, and Scott Klasky. 2019. Multilevel techniques for compression and reduction of scientific data—the multivariate case. SIAM Journal on Scientific Computing 41, 2 (2019), A1278–A1303

    work page 2019

  3. [3]

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. 2021. ViViT: A Video Vision Transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) . 6816–6826. doi:10.1109/ICCV48922.2021.00676

    arXiv 2021

  4. [4]

    Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick John- ston. 2018. Variational image compression with a scale hyperprior.arXiv preprint arXiv:1802.01436 (2018)

    arXiv 2018

  5. [5]

    Rafael Ballester-Ripoll, Peter Lindstrom, and Renato Pajarola. 2019. TTHRESH: Tensor compression for multidimensional visual data. IEEE transactions on visualization and computer graphics 26, 9 (2019), 2891–2903

    work page 2019

  6. [6]

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. 2021. Is space-time attention all you need for video understanding?. In ICML, Vol. 2. 4

    2021

  7. [7]

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. 2023. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)

    arXiv 2023

  8. [8]

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. 2023. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 22563–22575

    work page 2023

Show all 44 references
  1. [9]

    Jill M Boyce, Renaud Doré, Adrian Dziembowski, Julien Fleureau, Joel Jung, Bart Kroon, Basel Salahieh, Vinod Kumar Malamal Vadakital, and Lu Yu. 2021. MPEG immersive video coding standard. Proc. IEEE 109, 9 (2021), 1521–1536

    2021

  2. [10]

    Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah

  3. [11]

    Alyson Fox, James Diffenderfer, Jeffrey Hittinger, Geoffrey Sanders, and Peter Lindstrom. 2020. Stability analysis of inline ZFP compression for floating-point data in iterative methods. SIAM Journal on Scientific Computing 42, 5 (2020), A2701–A2730

    2020

  4. [12]

    Jean-Christophe Golaz, Peter M Caldwell, Luke P Van Roekel, Mark R Petersen, Qi Tang, Jonathan D Wolfe, Guta Abeshu, Valentine Anantharaj, Xylar S Asay-Davis, David C Bader, et al. 2019. The DOE E3SM coupled model version 1: Overview and evaluation at standard resolution. Jour...

    2019

  5. [13]

    Qian Gong, Jieyang Chen, Ben Whitney, Xin Liang, Viktor Reshniak, Tania Banerjee, Jaemoon Lee, Anand Rangarajan, Lipeng Wan, Nicolas Vidal, et al. 2023. MGARD: A multigrid framework for high-performance, error-controlled data compression and refactoring. SoftwareX 24 (2023), 101590

    2023

  6. [14]

    Chunming He, Yuqi Shen, Chengyu Fang, Fengyang Xiao, Longxiang Tang, Yulun Zhang, Wangmeng Zuo, Zhenhua Guo, and Xiu Li. 2025. Diffusion models in low-level vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    2025

  7. [15]

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. 2022. Video diffusion models. arXiv:2204.03458 (2022)

    2022 arXiv

  8. [16]

    Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. 2021. Variational diffusion models. Advances in neural information processing systems 34 (2021), 21696–21707

    2021

  9. [17]

    Diederik P Kingma and Max Welling. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013). SC’25, November 16–21, 2025, St. Louis, MO, USA Xiao Li, Liangji Zhu, Anand Rangarajan, Sanjay Ranka

    2013 arXiv

  10. [18]

    Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. 2022. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960 (2022)

    2022 arXiv

  11. [19]

    Jaemoon Lee, Qian Gong, Jong Choi, Tania Banerjee, Scott Klasky, Sanjay Ranka, and Anand Rangarajan. 2022. Error-Bounded Learned Scientific Data Compres- sion with Preservation of Derived Quantities. Applied Sciences 12, 13 (Jul 2022),

    2022

  12. [20]

    Jaemoon Lee, Xiao Li, Liangji Zhu, Sanjay Ranka, and Anand Rangarajan. 2025. Guaranteed Conditional Diffusion: 3D Block-based Models for Scientific Data Compression. arXiv preprint arXiv:2502.12951 (2025)

    2025 arXiv

  13. [21]

    Jaemoon Lee, Anand Rangarajan, and Sanjay Ranka. 2023. Nonlinear-by-Linear: Guaranteeing Error Bounds in Compressive Autoencoders(IC3-2023). Association for Computing Machinery, New York, NY, USA, 552–561. doi:10.1145/3607947. 3609702

    2023 doi

  14. [22]

    Xiao Li, Qian Gong, Jaemoon Lee, Scott Klasky, Anand Rangarajan, and Sanjay Ranka. 2024. Machine Learning Techniques for Data Reduction of Climate Applications. arXiv preprint arXiv:2405.00879 (2024)

    2024 arXiv

  15. [23]

    Xiao Li, Jaemoon Lee, Anand Rangarajan, and Sanjay Ranka. 2024. Attention based machine learning methods for data reduction with guaranteed error bounds. In 2024 IEEE International Conference on Big Data (BigData) . IEEE, 1039–1048

    2024

  16. [24]

    Xiao Li, Jaemoon Lee, Anand Rangarajan, and Sanjay Ranka. 2024. Attention Based Machine Learning Methods for Data Reduction with Guaranteed Error Bounds. arXiv preprint arXiv:2409.05357 (2024)

    2024 arXiv

  17. [25]

    Xiao Li, Jaemoon Lee, Anand Rangarajan, and Sanjay Ranka. 2024. Foundation Model for Lossy Compression of Spatiotemporal Scientific Data. arXiv preprint arXiv:2412.17184 (2024)

    2024 arXiv

  18. [26]

    Liang, S

    X. Liang, S. Di, D. Tao, S. Li, S. Li, H. Guo, Z. Chen, and F. Cappello. 2018. Error- controlled lossy compression optimized for high compression ratios of scientific datasets. In 2018 IEEE International Conference on Big Data (Big Data) . 438–447. doi:10.1109/BigData.2018.8622520

    2018

  19. [27]

    Xin Liang, Kai Zhao, Sheng Di, Sihuan Li, Robert Underwood, Ali M Gok, Jiannan Tian, Junjing Deng, Jon C Calhoun, Dingwen Tao, et al. 2022. SZ3: A modular framework for composing prediction-based error-bounded lossy compressors. IEEE Transactions on Big Data 9, 2 (2022), 485–498

    2022

  20. [28]

    Peter Lindstrom. 2014. Fixed-Rate Compressed Floating-Point Arrays. IEEE Transactions on Visualization and Computer Graphics 20 (08 2014). doi:10.1109/ TVCG.2014.2346458

    2014

  21. [29]

    Jinyang Liu, Sheng Di, Kai Zhao, Xin Liang, Zizhong Chen, and Franck Cappello

  22. [30]

    David Minnen, Johannes Ballé, and George D Toderici. 2018. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems 31 (2018)

    2018

  23. [31]

    Sungkwang Mun and James E Fowler. 2012. DPCM for quantized block-based compressed sensing of images. In 2012 Proceedings of the 20th European Signal Processing Conference (EUSIPCO). IEEE, 1424–1428

    2012

  24. [32]

    In Proceedings of the 37th International Conference on Supercomputing

    Faz: A flexible auto-tuned modular error-bounded compression frame- work for scientific data. In Proceedings of the 37th International Conference on Supercomputing. 1–13

  25. [33]

    Jorma Rissanen and Glen Langdon. 1981. Universal modeling and coding. IEEE Transactions on Information Theory 27, 1 (1981), 12–23

    1981

  26. [34]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

    2022

  27. [35]

    Lucas Relic, Roberto Azevedo, Markus Gross, and Christopher Schroers. 2024. Lossy image compression with foundation diffusion models. In European Confer- ence on Computer Vision . Springer, 303–319

    2024

  28. [36]

    Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. 2022. MCVD: Masked Conditional Video Diffusion for Prediction, Generation, and Interpolation. In (NeurIPS) Advances in Neural Information Processing Systems . https://arxiv. org/abs/2205.09853

    2022 arXiv

  29. [37]

    M Wan, S Chen, G Eyink, C Meneveau, E Perlman, R Burns, Y Li, A Szalay, and S Hamilton. 2016. Johns Hopkins Turbulence Database (JHTDB)

    2016

  30. [38]

    Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. 2016. Conditional image generation with pixelcnn decoders. Ad- vances in neural information processing systems 29 (2016)

    2016

  31. [39]

    Chun Sang Yoo, Tianfeng Lu, Jacqueline H Chen, and Chung K Law. 2011. Direct numerical simulations of ignition of a lean n-heptane/air mixture with temper- ature inhomogeneities at constant volume: Parametric study. Combustion and Flame 158, 9 (2011), 1727–1741

    2011

  32. [40]

    C. S. Yoo, T. Lu, J. H. Chen, and C. K. Law. 2011. Direct numerical simulations of ignition of a lean𝑛−heptane/air mixture with temperature inhomogeneities at constant volume: Parametric study. Combust. Flame 158 (2011), 1727–1741

    2011

  33. [41]

    Ruihan Yang and Stephan Mandt. 2023. Lossy image compression with conditional diffusion models. Advances in Neural Information Processing Systems 36 (2023), 64971–64995

    2023

  34. [44]

    Zheyuan Zhan, Defang Chen, Jian-Ping Mei, Zhenghe Zhao, Jiawei Chen, Chun Chen, Siwei Lyu, and Can Wang. 2024. Conditional Image Synthesis with Diffu- sion Models: A Survey. arXiv preprint arXiv:2409.19365 (2024)

    2024 arXiv

  35. [2023]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 9 (2023), 10850–10869

    Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 9 (2023), 10850–10869

    2023

  36. [6718]

    doi:10.3390/app12136718

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.