Pith. sign in

REVIEW 41 references

CSI Prediction Using Diffusion Models

T0 review · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Diffusion-based CSI predictors, which generate future channel samples from a temporal encoder's latent code, outperform deterministic deep-learning baselines on 3GPP CDL channels, with NMSE gains up to 5–8 dB at high SNR.

desk verdict A genuinely new application of diffusion to CSI prediction with a plausible but not yet pinned-down 5–8 dB NMSE claim; the single-sample evaluation under NMSE is the core weakness. read the letter →

arxiv 2510.11214 v2 pith:I7EMAMYW submitted 2025-10-13 eess.SP cs.ITmath.IT

Mehdi Sattari , Javad Aliakbari , Alexandre Graell i Amat , Tommy Svensson This is my paper · ORCID
classification eess.SPcs.ITmath.IT
keywords CSIpredictiondiffusionmodelsMIMOchannelstateinformationgenerativemodelingDDIMtemporalencoder3GPPCDL
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 sets out to show that CSI prediction—forecasting a MIMO channel a few OFDM symbols ahead—is better posed as a probabilistic generation problem than as deterministic regression. It decomposes the task into a temporal encoder that compresses observed channel history into latent features, and a diffusion generator that denoises random noise into future CSI conditioned on those features. On simulated 3GPP CDL channels at 28 GHz, the authors report that this diffusion framework consistently beats GRU, ConvLSTM, and LinFormer baselines, with the U-Net-based autoregressive variant achieving 5–8 dB NMSE gains at short horizons and high SNR, and that three DDIM sampling steps suffice near-100-step quality. They further report stronger generalization to a 3 GHz distribution shift than the deterministic baselines. If correct, the claim matters because it would make generative models a practical option for a latency-sensitive, low-overhead task previously dominated by deterministic RNN/Transformer predictors.

What carries the argument

The central object is a conditional diffusion generator coupled to a temporal encoder. The temporal encoder (ConvLSTM or LinFormer) maps the observed history into a latent representation Z; the generator (U-Net, DiT, or 3D U-Net) learns the reverse diffusion transitions p_theta(H_{t-1} | H_t, Z), so the whole system approximates the conditional distribution p(H_f | H_p) instead of a point mapping. DDIM scheduling with a small number of steps makes sampling fast enough for real-time use, and the same training loss supports both autoregressive and seq2seq inference.

What would settle it

A reader could take the trained DiU model and compare NMSE for a single sample versus the average of 10–100 samples; if averaging does not improve (or does not beat a deterministic U-Net with the same temporal encoder and loss), the 'probabilistic outperformance' claim is not established. A second check: rerun the comparison with a deterministic U-Net generator (same architecture, same conditioning) trained with the same Huber loss and see whether the 5–8 dB gain survives.

Watch

Extended reading notes

Core claim

The paper's central claim: modeling the conditional distribution p(H_f | H_p) of future CSI given past CSI, rather than fitting a deterministic mapping, yields higher NMSE accuracy and better robustness to domain shift. The predictor is a conditional diffusion model: a temporal encoder (ConvLSTM or LinFormer) compresses the observed history into a latent code, and a diffusion generator denoises a noisy future frame conditioned on that code, learning the reverse process over a 2000-step squared-cosine noise schedule. At inference, DDIM runs the reverse chain from Gaussian noise to draw one future-frame sample. Autoregressive and seq2seq modes, plus U-Net, DiT, and 3D U-Net backbones, are eval

Load-bearing premise

The reported NMSE superiority assumes that a single stochastic sample drawn from the learned conditional distribution is competitive with the conditional-mean estimate that NMSE mathematically rewards, without averaging over samples or isolating the diffusion mechanism from the U-Net backbone.

Editorial extensions

If this is right

  • A single trained diffusion predictor can run in either autoregressive or seq2seq mode with the same training objective, giving deployment flexibility across contexts and horizons.
  • DDIM with as few as three sampling steps achieves near-100-step accuracy, so diffusion-based CSI prediction becomes feasible for latency-constrained links.
  • The framework accommodates different temporal encoders and diffusion backbones, letting system designers trade compute (FLOPs and parameters) against NMSE.
  • Diffusion predictors trained at 28 GHz degrade less when evaluated at 3 GHz, suggesting they learn channel structure rather than frequency-specific artifacts.
  • Seq2seq diffusion predicts the whole horizon in one pass, avoiding error accumulation from recursive prediction, at the cost of fixed context length and horizon.

Reading between the lines

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

  • Because the paper scores NMSE with a single stochastic diffusion sample, averaging many samples at inference is a natural next experiment; if the average improves NMSE, the reported numbers understate the model, and if it does not, the gains may come from the U-Net architecture rather than the probabilistic formulation.
  • The framework calls for distributional evaluation—prediction intervals, CRPS, or coverage—where the uncertainty captured by stochastic sampling can be directly tested, something NMSE alone cannot reveal.
  • The unified no-temporal-encoder variant points to a minimal-cost deployment path; a testable extension is distilling the conditional mean from the diffusion model and comparing it against the strongest deterministic baseline to isolate the value of the generative formulation.
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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central claim is an empirical benchmark comparison, and no prediction reduces by construction to a fitted input or a self-citation.

full rationale

The paper's central claim is an empirical NMSE comparison of diffusion-based CSI predictors against deterministic baselines, supported by simulations on a 3GPP CDL dataset. There is no analytical chain in which an output is equivalent by construction to an input. The MMSE/conditional-mean discussion in Eqs. (20)-(21) is used only as motivation for decomposing the predictor into a temporal encoder and a generator; the paper does not claim that a single diffusion sample equals the conditional mean, nor does it relabel a fitted parameter as a prediction. The DDIM scheduler, squared-cosine noise schedule, Huber loss, and network backbones are either standard external components or explicit implementation choices; none is justified solely by a load-bearing self-citation, and the cited works (e.g., [32]-[36], [39]-[41]) are independent external sources. The generalization experiment in Fig. 7 does use a test set generated by the same CDL simulator pipeline, but at a different carrier frequency (3 GHz vs. 28 GHz), so it is a held-out distribution-shift evaluation rather than a circular reuse of training data. The potential concern that NMSE is optimized by the conditional mean while Algorithms 2 and 3 return a single stochastic sample is a methodological/evaluation issue, not a circularity issue: nothing in the paper defines the reported NMSE result as the fitted optimum or derives it from Eq. (21). No quoted passage exhibits the required specific reduction, so under the stated rules the correct finding is no significant circularity.

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

The central claim is empirical; its load-bearing inputs are simulation choices, hyperparameters, and an unstated metric assumption. The model introduces no new physical or mathematical entities. The main ad hoc assumption is that a single random diffusion sample can compete under NMSE, which is not justified.

free parameters (5)
  • Training SNR distribution rho = rho in [-20,20] dB (uniform)
    Algorithm 1 samples rho to corrupt past CSI; the text says dB but the algorithm uses sqrt(rho) as a linear gain, an inconsistency that affects all reported SNR results.
  • Huber loss threshold delta = 0.016
    Chosen without a sensitivity analysis; it changes the loss landscape and therefore the NMSE ranking across models.
  • Diffusion schedule T and clipping = T=2000, beta_min=1e-4, beta_max=2e-2
    Adopted from the squared-cosine schedule of [41]; no ablation of the schedule is provided.
  • Dataset composition = 100k samples; Np=30, Nf=10; velocity 30-120 km/h; delay spread 50-400 ns; CDL-A to CDL-E
    These simulation choices define the benchmark, and the paper provides no evaluation on measured channels or an external dataset.
  • Inference sampling steps = 3-100 DDIM steps
    The paper claims 3 steps suffice, but the accuracy-complexity tradeoff is shown only in a single curve at 10 dB SNR.
assumptions (5)
  • domain assumption 3GPP CDL channel model with the specified parameter ranges adequately represents the CSI prediction problem.
    All training and evaluation data are generated from Eqs. (1)-(4); no real channel measurements are used.
  • standard math DDPM/DDIM theory (Eqs. 6-19) holds as background.
    Standard diffusion-model background is used without modification.
  • ad hoc to paper A single stochastic sample from the learned conditional distribution is competitive under NMSE, despite Eq. (21) identifying the conditional mean as NMSE-optimal.
    Algorithms 2 and 3 output one random sample with no averaging; the gap between stochastic sampling and the NMSE-optimal CME is never addressed.
  • domain assumption Min-max scaling to (0,1) preserves the prediction problem and the NMSE metric.
    Applied to the whole dataset; its effect on a ratio-of-norms metric is not analyzed.
  • domain assumption The 100,000 independently generated CSI samples provide an unbiased estimate of temporal channel dynamics without leakage.
    No train/test split or temporal correlation analysis is reported; each sample is treated as independent even though real channels have long time correlation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CSI Prediction Using Diffusion Models." pith.science (2026). https://pith.science/paper/I7EMAMYW

@misc{pith2026251011214,
  author       = {Pith},
  title        = {Pith review of: CSI Prediction Using Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I7EMAMYW}},
  note         = {Machine review of arXiv:2510.11214}
}
read the original abstract

Acquiring accurate channel state information (CSI) is critical for reliable and efficient wireless communication, but challenges such as high pilot overhead and channel aging hinder timely and accurate CSI acquisition. CSI prediction, which forecasts future CSI from historical observations, offers a promising solution. Recent deep learning approaches, including recurrent neural networks and Transformers, have achieved notable success but typically learn deterministic mappings, limiting their ability to capture the stochastic and multimodal nature of wireless channels. In this paper, we introduce a novel probabilistic framework for CSI prediction based on diffusion models, offering a flexible design that supports integration of diverse prediction schemes. We decompose the CSI prediction task into two components: a temporal encoder, which extracts channel dynamics, and a diffusion-based generator, which produces future CSI samples. We investigate two inference schemes-autoregressive and sequence-to-sequence- and explore multiple diffusion backbones, including U-Net and Transformer-based architectures. Furthermore, we examine a diffusion-based approach without an explicit temporal encoder and utilize the DDIM scheduling to reduce model complexity. Extensive simulations demonstrate that our diffusion-based models significantly outperform state-of-the-art baselines.

Figures

Figures reproduced from arXiv: 2510.11214 by the authors.

Figure 1
Figure 1. CSI prediction with a diffusion model conditioned [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. NMSE of different CSI prediction models over varying prediction steps at inference SNRs of [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. NMSE performance of CSI prediction models versus inference SNR. Results are reported for the first prediction step, [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: NMSE performance of CSI prediction models as a function of prediction step, evaluated on CSI samples with user [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: NMSE performance versus prediction step at [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Left: NMSE performance versus prediction step at [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Left: NMSE performance versus prediction step, evaluated at [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 10 linked inside Pith

  1. [1]

    Biglieri, R

    E. Biglieri, R. Calderbank, A. Constantinides, A. Goldsmith, A. Paulraj, and H. V . Poor,MIMO Wireless Communications. Cambridge, U.K.: Cambridge Univ. Press, 2007

    2007

  2. [2]

    How much training is needed in multiple-antenna wireless links?

    B. Hassibi and B. M. Hochwald, “How much training is needed in multiple-antenna wireless links?”IEEE Trans. Inf. Theory, vol. 49, no. 4, pp. 951–963, 2003

    2003

  3. [3]

    Five disruptive technology directions for 5G,

    F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski, “Five disruptive technology directions for 5G,”IEEE Commun. Mag., vol. 52, no. 2, pp. 74–80, 2014

    2014

  4. [4]

    Channel estimation and hybrid precoding for millimeter wave cellular systems,

    A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, “Channel estimation and hybrid precoding for millimeter wave cellular systems,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831–846, 2014

    2014

  5. [5]

    Towards systems beyond 3G based on adaptive OFDMA transmission,

    M. Sternad, T. Svensson, T. Ottosson, A. Ahl ´en, A. Svensson, and A. Brunstrom, “Towards systems beyond 3G based on adaptive OFDMA transmission,”Proc. IEEE, vol. 95, no. 12, pp. 2432–2455, 2007

    2007

  6. [6]

    Addressing the curse of mobility in massive MIMO with prony-based angular-delay domain channel predictions,

    H. Yin, H. Wang, Y . Liu, and D. Gesbert, “Addressing the curse of mobility in massive MIMO with prony-based angular-delay domain channel predictions,”IEEE J. Sel. Areas Commun., vol. 38, no. 12, pp. 2903–2917, 2020

    2020

  7. [7]

    Massive MIMO channel prediction: Kalman filtering vs. machine learning,

    H. Kim, S. Kim, H. Lee, C. Jang, Y . Choi, and J. Choi, “Massive MIMO channel prediction: Kalman filtering vs. machine learning,”IEEE Trans. Commun., vol. 69, no. 1, pp. 518–528, 2021

    2021

  8. [8]

    Joint channel estimation and prediction for OFDM systems,

    I. C. Wong and B. L. Evans, “Joint channel estimation and prediction for OFDM systems,” inProc. IEEE Global Telecommun. Conf. (GLOBE- COM), 2005, pp. 5 pp.–2259

    2005

Show all 41 references
  1. [9]

    Spatial wireless channel prediction under location uncertainty,

    L. S. Muppirisetty, T. Svensson, and H. Wymeersch, “Spatial wireless channel prediction under location uncertainty,”IEEE Trans. Wireless Commun., vol. 15, no. 2, pp. 1031–1044, 2016

    2016

  2. [10]

    Predictor antenna: A technique to boost the performance of moving relays,

    H. Guo, B. Makki, D.-T. Phan-Huy, E. Dahlman, M.-S. Alouini, and T. Svensson, “Predictor antenna: A technique to boost the performance of moving relays,”IEEE Commun. Mag., vol. 59, no. 7, pp. 80–86, 2021

    2021

  3. [11]

    Autoregressive modeling for fading channel simulation,

    K. E. Baddour and N. C. Beaulieu, “Autoregressive modeling for fading channel simulation,”IEEE Trans. Wireless Commun., vol. 4, no. 4, pp. 1650–1662, 2005

    2005

  4. [12]

    Deep learning with long short-term memory networks for financial market predictions,

    T. Fischer and C. Krauss, “Deep learning with long short-term memory networks for financial market predictions,”Eur. J. Oper. Res., vol. 270, no. 2, pp. 654–669, 2018

    2018

  5. [13]

    Short-term residential load forecasting based on LSTM recurrent neural network,

    W. Kong, Z. Y . Dong, Y . Jia, D. J. Hill, Y . Xu, and Y . Zhang, “Short-term residential load forecasting based on LSTM recurrent neural network,” IEEE Trans. Smart Grid, vol. 10, no. 1, pp. 841–851, 2019

    2019

  6. [14]

    Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting,

    B. Yu, H. Yin, and Z. Zhu, “Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting,” inProc. Int. Joint Conf. Artif. Intell. (IJCAI), 2018, pp. 3634–3640

    2018

  7. [15]

    The rise of diffusion models in time-series forecasting,

    C. Meijer and L. Y . Chen, “The rise of diffusion models in time-series forecasting,” 2024, arXiv:2401.03006

    2024 arXiv

  8. [16]

    Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting,

    K. Rasul, C. Seward, I. Schuster, and R. V ollgraf, “Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting,” 2021, arXiv:2101.12072

    2021 arXiv

  9. [17]

    Channel prediction in high-mobility massive MIMO: From spatio-temporal autoregression to deep learning,

    C. Wu, X. Yi, Y . Zhu, W. Wang, L. You, and X. Gao, “Channel prediction in high-mobility massive MIMO: From spatio-temporal autoregression to deep learning,”IEEE J. Sel. Areas Commun., vol. 39, no. 7, pp. 1915–1930, 2021

    1915

  10. [18]

    A comparison of neural networks for wireless channel prediction,

    O. Stenhammar, G. Fodor, and C. Fischione, “A comparison of neural networks for wireless channel prediction,”IEEE Wireless Commun., vol. 31, no. 3, pp. 235–241, 2024

    2024

  11. [19]

    Spatio- temporal neural network for channel prediction in massive MIMO- OFDM systems,

    G. Liu, Z. Hu, L. Wang, J. Xue, H. Yin, and D. Gesbert, “Spatio- temporal neural network for channel prediction in massive MIMO- OFDM systems,”IEEE Trans. Commun., vol. 70, no. 12, pp. 8003–8016, 2022

    2022

  12. [20]

    Recurrent neural network-based frequency-domain channel prediction for wideband communications,

    W. Jiang and H. D. Schotten, “Recurrent neural network-based frequency-domain channel prediction for wideband communications,” inProc. IEEE Veh. Technol. Conf. (VTC Spring), 2019, pp. 1–6

    2019

  13. [21]

    Deep learning for fading channel prediction,

    ——, “Deep learning for fading channel prediction,”IEEE Open J. Commun. Soc., vol. 1, pp. 320–332, 2020

    2020

  14. [22]

    CSI-LLM: A novel downlink channel prediction method aligned with LLM pre-training,

    S. Fan, Z. Liu, X. Gu, and H. Li, “CSI-LLM: A novel downlink channel prediction method aligned with LLM pre-training,” inProc. IEEE Wireless Commun. Netw. Conf. (WCNC), 2025, pp. 1–6

    2025

  15. [23]

    Machine learning for fu- ture wireless communications: Channel prediction perspectives,

    H. Kim, J. Choi, and D. J. Love, “Machine learning for fu- ture wireless communications: Channel prediction perspectives,” 2025, arXiv:2502.18196

    2025 arXiv

  16. [24]

    Time- varying channel prediction for RIS-assisted MU-MISO networks via deep learning,

    W. Xu, J. An, Y . Xu, C. Huang, L. Gan, and C. Yuen, “Time- varying channel prediction for RIS-assisted MU-MISO networks via deep learning,”IEEE Trans. Cogn. Commun. Netw., vol. 8, no. 4, pp. 1802–1815, 2022

    2022

  17. [25]

    Accurate channel prediction based on transformer: Making mobility negligible,

    H. Jiang, M. Cui, D. W. K. Ng, and L. Dai, “Accurate channel prediction based on transformer: Making mobility negligible,”IEEE J. Sel. Areas Commun., vol. 40, no. 9, pp. 2717–2732, 2022

    2022

  18. [26]

    Linformer: A linear-based lightweight transformer architecture for time-aware MIMO channel prediction,

    Y . Jin, Y . Wu, Y . Gao, S. Zhang, S. Xu, and C.-X. Wang, “Linformer: A linear-based lightweight transformer architecture for time-aware MIMO channel prediction,”IEEE Trans. Wireless Commun., pp. 1–1, 2025

    2025

  19. [27]

    Enhancing reliability in AI-based CSI prediction: A proxy-based performance monitoring ap- proach,

    C. Jiang, J. Guo, C.-K. Wen, and S. Jin, “Enhancing reliability in AI-based CSI prediction: A proxy-based performance monitoring ap- proach,”IEEE Trans. Commun., vol. 73, no. 4, pp. 2602–2615, 2025

    2025

  20. [28]

    Next-slot OFDM- CSI prediction: Multi-head self-attention or state space model?

    M. Akrout, F. Bellili, A. Mezghani, and R. W. Heath, “Next-slot OFDM- CSI prediction: Multi-head self-attention or state space model?” 2024, arXiv:2405.11072

    2024 arXiv

  21. [29]

    Reverse ordering techniques for attention-based channel prediction,

    V . Rizzello, B. B ¨ock, M. Joham, and W. Utschick, “Reverse ordering techniques for attention-based channel prediction,”IEEE Open J. Signal Process., vol. 5, pp. 248–256, 2024

    2024

  22. [30]

    CSI-BERT2: A BERT-inspired framework for efficient CSI prediction and classification in wireless communication and sensing,

    Z. Zhao, F. Meng, Z. Lyu, H. Li, X. Li, and G. Zhu, “CSI-BERT2: A BERT-inspired framework for efficient CSI prediction and classification in wireless communication and sensing,” 2025, arXiv:2412.06861

    2025

  23. [31]

    Spectral temporal graph neural network for massive MIMO CSI prediction,

    S. Mourya, P. Reddy, S. Amuru, and K. K. Kuchi, “Spectral temporal graph neural network for massive MIMO CSI prediction,”IEEE Wireless Commun. Lett., vol. 13, no. 5, pp. 1399–1403, 2024

    2024

  24. [32]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inAdv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, 2020, pp. 6840– 6851

    2020

  25. [33]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” 2022, arXiv:2010.02502

    2022 arXiv

  26. [34]

    U-Net: Convolutional net- works for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional net- works for biomedical image segmentation,” 2015, arXiv:1505.04597

    2015 arXiv

  27. [35]

    3D U-Net: Learning dense volumetric segmentation from sparse anno- tation,

    ¨O. C ¸ ic ¸ek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and O. Ronneberger, “3D U-Net: Learning dense volumetric segmentation from sparse anno- tation,” inProc. Int. Conf. Med. Image Comput. Comput.-Assist. Interv. (MICCAI), 2016, pp. 424–432

    2016

  28. [36]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” 2023, arXiv:2212.09748

    2023 arXiv

  29. [37]

    Convo- lutional LSTM network: A machine learning approach for precipitation nowcasting,

    X. Shi, Z. Chen, H. Wang, D.-Y . Yeung, W. Wong, and W. Woo, “Convo- lutional LSTM network: A machine learning approach for precipitation nowcasting,” 2015, arXiv:1506.04214

    2015 arXiv

  30. [38]

    Study on channel model for frequencies from 0.5 to 100 GHz,

    “Study on channel model for frequencies from 0.5 to 100 GHz,” 3GPP, Tech. Rep. TR 38.901, Mar. 2022, available: https://www.3gpp.org/ftp/ Specs/archive/38 series/38.901/

    2022

  31. [39]

    On the asymptotic mean square error optimality of diffusion models,

    B. Fesl, B. B ¨ock, F. Strasser, M. Baur, M. Joham, and W. Utschick, “On the asymptotic mean square error optimality of diffusion models,” 2025, arXiv:2403.02957

    2025 arXiv

  32. [40]

    Robust estimation of a location parameter,

    P. J. Huber, “Robust estimation of a location parameter,”Ann. Math. Statist., vol. 35, no. 1, pp. 73–101, 1964

    1964

  33. [41]

    Improved denoising diffusion probabilistic models,

    A. Nichol and P. Dhariwal, “Improved denoising diffusion probabilistic models,” 2021, arXiv:2102.09672

    2021 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.