Pith. sign in

REVIEW 4 major objections 5 minor 48 references

DIFFUMA: High-Fidelity Spatio-Temporal Video Prediction via Dual-Path Mamba and Diffusion Enhancement

T0 review · 4 major / 5 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 DIFFUMA couples bidirectional Mamba temporal modeling with a diffusion detail enhancer to predict sharp future frames, backed by a new wafer-dicing video benchmark.

desk verdict The paper's headline CHDL numbers don't match its own tables, and the dataset's 'temporal' axis is actually spatial translation, which guts the central claim. read the letter →

arxiv 2507.06738 v1 pith:CRN7A6S3 submitted 2025-07-09 cs.CV cs.AI

Xinyu Xie , Weifeng Cao , Jun Shi , Yangyang Hu , Hui Liang , Wanyong Liang , Xiaoliang Qian This is my paper · ORCID
classification cs.CVcs.AI
keywords Spatio-temporalvideopredictionChipDicingLaneDatasetIndustrialAIMambastatespacemodelDiffusionenhancementSemiconductorwaferDual-patharchitecture
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 a single architecture can solve both halves of high-fidelity video prediction: long-range temporal dynamics and fine-grained spatial detail. To test this in an industrial setting, the authors release CHDL, which they describe as the first public temporal-image dataset of the semiconductor wafer dicing process. On CHDL they report that their model, DIFFUMA, outperforms existing predictors by a large margin, and they also report strong results on the WeatherBench suite. The reason the claim matters is that chip-dicing monitoring, defect detection, and digital twins all need predictions that stay sharp instead of blurring over time.

What carries the argument

The load-bearing machinery is the dual-path architecture: a Mamba Module that models temporal dynamics through a spatial encoder, stacked bidirectional Mamba blocks with forward and backward state-space scans and gated fusion, and a spatial decoder; and a Diffusion Module built from 12 DiT-style blocks with multi-head self-attention and MLPs, modulated by timestep and Mamba-derived context through adaptive layer normalization. The Diffusion Module is trained to predict the noise added to the input, and at inference, with timestep zero, its output becomes a detail-enhancement residual added to the Mamba prediction. The conditioning of the denoiser on the Mamba module is what lets one path guide the other.

What would settle it

Take random CHDL samples and predict their target frames with simple nearest-neighbor or bilinear interpolation between the surrounding acquisition positions; if such a non-learned baseline closes most of the gap to DIFFUMA's reported CHDL SSIM of 0.9254 and MSE of 0.0371, the prediction task is mostly spatial interpolation. Additionally, split CHDL so that target frames never come from the same dicing lane as any training frame and rerun the comparison, which would separate true generalization from memorization of repeated lane textures.

Watch

Extended reading notes

Core claim

DIFFUMA splits the prediction task into a temporal path and a spatial path. A bidirectional Mamba module reads the frame sequence, compresses each frame with a convolutional encoder, runs forward and backward state-space scans, and gates their fusion to produce a coherent motion forecast. In parallel, a diffusion-inspired module takes a noised version of the input, injects both a timestep embedding and the Mamba forecast as conditions through adaptive normalization, and learns to predict the noise in a single pass; at inference the module's output acts as a residual that sharpens the Mamba prediction. The two paths are trained jointly with a denoising loss plus an L1 frame-reconstruction loss. The accompanying claim is that CHDL is the first public video benchmark of the semiconductor dicing process, captured at 800x600 grayscale along dicing lanes, with 10-frame samples split 5 input and 5 target.

Load-bearing premise

The load-bearing premise is that CHDL's 10 frames per sample form a temporal sequence in which each future frame is causally determined by earlier frames; if instead the frames are just adjacent patches of a nearly periodic lane texture photographed at different stage positions, the benchmark tests texture extrapolation rather than video forecasting.

Editorial extensions

If this is right

  • If the reported gains hold, DIFFUMA would set a new state of the art on CHDL and on WeatherBench, with especially large improvements over SimVP.
  • A single-pass diffusion-style denoiser conditioned on temporal features can restore texture detail without iterative sampling, which keeps inference affordable for industrial monitoring.
  • CHDL would give the community a public benchmark for process forecasting, micro-defect detection, and digital-twin simulation in semiconductor manufacturing.
  • The dual-path design suggests that future predictors can combine linear-complexity sequence models with generative detail enhancement rather than choosing one paradigm.

Reading between the lines

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

  • The abstract's claim of a 39% MSE reduction on CHDL and an SSIM of 0.988 does not match Table 2, which reports a CHDL SSIM of 0.9254 and an MSE reduction of about 30% relative to SimVP; reconciling this discrepancy is a natural first check before taking the headline numbers at face value.
  • Because CHDL frames are captured at equidistant stage positions rather than over elapsed time, a telling test is whether simple non-learned baselines, such as copying the last frame or interpolating between neighboring positions, already achieve most of the reported accuracy; if so, the benchmark may measure texture extrapolation rather than video forecasting.
  • The same conditioning idea could be carried to other fine-grained industrial processes, such as etching or laser drilling, or to extreme-weather forecasting, where preserving fine structure is the primary bottleneck.
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 / 5 minor

Summary. The paper introduces CHDL, an industrial imaging dataset of semiconductor wafer dicing lanes, and DIFFUMA, a dual-path video prediction model that combines a bidirectional Mamba temporal backbone with a diffusion-style detail-enhancement module. The authors report state-of-the-art results on CHDL and WeatherBench, claiming a 39% MSE reduction and an SSIM improvement to 0.988 on CHDL. The manuscript also proposes CHDL as the first public temporal video benchmark for the semiconductor dicing process. However, the paper contains multiple internal inconsistencies in the headline quantitative claims, the dataset statistics, and the prediction configuration, and the acquisition protocol raises a fundamental question about whether the CHDL frames are actually temporally ordered. The central empirical and methodological claims are therefore not currently supported by the manuscript's own evidence.

Significance. If the dataset were a valid temporal video benchmark and the reported results were reproducible, this paper would provide a useful industrial resource and an interesting architectural contribution: the combination of a Mamba-based temporal encoder with a diffusion-style detail enhancer is a reasonable idea, and the WeatherBench results show consistent improvements over the tested baselines. However, the significance is severely undermined by three load-bearing problems. First, the headline CHDL numbers in the abstract and introduction are contradicted by the paper's own Table 2. Second, the dataset statistics and the input/output lengths differ between Section 3.2 and Table 1, making the experimental setup and dataset contribution unverifiable. Third, the acquisition procedure described in Section 3.1 indicates that consecutive frames are captured at different spatial positions along a dicing lane rather than at the same location over time, which would make the task spatial texture extrapolation rather than spatiotemporal forecasting.

major comments (4)
  1. [Abstract, Section 1, and Table 2] The abstract and Section 1 state that on CHDL, DIFFUMA reduces MSE by 39% and improves SSIM from 0.926 to 0.988. Table 2 reports CHDL SSE for DiffuMa as 0.9254 and for SimVP as 0.8350, which corresponds to an SSIM improvement of about 0.09, not 0.06, and an MSE reduction of about 29.7% (0.0371 vs 0.0528), not 39%. Additionally, Section 5.3 states that the SSIM reaches 0.988 on WeatherBench and 0.926±0.018 on CHDL, again contradicting the abstract's attribution of 0.988 to CHDL. Since the abstract quantifies the central contribution, this inconsistency is load-bearing for the paper's headline claim.
  2. [Section 3.2 and Table 1] Section 3.2 states that the final dataset comprises 6,000 input-target sample pairs, evenly split into 3,000 for training and 3,000 for validation. Table 1 reports Ntrain=4,874 and Ntest=1,216, which is inconsistent with both the total count and the split. The same section defines the input as the first 5 frames and the target as the subsequent 5 frames, whereas Table 1 reports T=5 and K=1, i.e., a one-frame prediction length. This makes the dataset statistics and the exact prediction task ambiguous, and the discrepancy is not explained anywhere in the paper.
  3. [Section 3.1 and Figure 2] The acquisition protocol in Section 3.1 says the automated alignment system moves along the dicing lanes at fixed intervals, acquiring images at 10 equidistant positions during wafer translation, and Figure 2 labels the horizontal axis as an isometric translation. Under this description, the 10 frames in each sample are recorded at different spatial locations along the lane, not at the same location over time. The prediction task is therefore extrapolation of a spatial texture, not forecasting of temporal dynamics. This directly undermines the paper's claim to introduce the first temporal video benchmark for the dicing process and calls into question whether the CHDL results demonstrate spatio-temporal video prediction at all.
  4. [Section 4.3, Eq. (8), and Eq. (7)] The diffusion module is trained to predict the noise added to the input frames X1:Tin (Eq. 8), and at inference its output is added as a residual to the Mamba prediction for the future frames (Eq. 7). There is no term in the loss that trains the diffusion module to denoise or enhance the target frames; the reconstruction loss in Eq. (9) supervises the final fused prediction, but the module that produces the residual is never directly trained to restore the future frames. Unless the input and target frames are essentially identical, which would trivialize the prediction task, the proposed denoising objective does not support the claim that the diffusion module improves the fidelity of the predicted future frames. This is a load-bearing gap in the method's design.
minor comments (5)
  1. [Data availability statement] The paper repeatedly calls CHDL the first public dataset for this domain, but the Data availability section states "Data will be made available on request," which does not meet the usual expectation of a public benchmark release.
  2. [Figure 7] Figure 7 includes a model labeled "MambaPastNet" that is not defined, described, or mentioned in the experimental setup; this appears to be either a typo or an unlisted baseline.
  3. [Naming throughout] The model name is used inconsistently, e.g., "DiffuMa" in the title and Table 2, "DIFFUMA" in the abstract, and "Mamba" as a standalone label in Figure 7; please standardize.
  4. [Section 5.2] The claim that DiffuMa reduces WeatherBench MSE by 39% and MAE by 21% on average relative to SimVP is not directly supported by Table 2, where the reduction varies by variable (for example, MSE reduction is about 51% for Cloud_cover but about 6% for Temperature); reporting a single averaged percentage without the averaging procedure is unclear.
  5. [Section 3.2] The text describes the dataset as split into "training" and "validation," while Table 1 uses "Ntest"; please clarify the role of the validation set and how the reported test metrics were obtained.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical claims rest on standard supervised losses and external baselines; internal inconsistencies are correctness/reproducibility concerns, not circular reasoning.

full rationale

The paper's central derivation chain is not circular. The final prediction is trained with a direct reconstruction loss (Eq. 9, L_recon = E[||Ŷ − Y||_1]) against the ground-truth future frames, so the reported MSE/MAE/SSIM numbers are not fitted inputs renamed as predictions. The auxiliary diffusion loss (Eqs. 1 and 8) trains a denoiser on noised input frames, but the end-to-end system is also explicitly supervised toward the target sequence; no evaluation metric is defined in terms of the model's own outputs. There are no load-bearing self-citations: the cited Mamba, SimVP, PastNet, and WeatherBench results are external works, and no uniqueness theorem or ansatz is imported from the authors' prior publications. Several issues in the paper are real but belong to correctness and reproducibility rather than circularity: (1) Section 3.1 describes CHDL frames as captured at 10 equidistant positions during wafer translation, so the 'temporal' axis may actually be spatial, which threatens the dataset's construct validity as a video benchmark; (2) the abstract's claim of 39% MSE reduction and SSIM 0.988 on CHDL conflicts with Table 2, which reports CHDL SSIM 0.9254 and a roughly 29.7% MSE reduction versus SimVP; (3) Table 1 lists CHDL as T=5, K=1 while Section 3.2 states 5-frame targets; and (4) the dataset is only 'available on request' despite being described as released. These inconsistencies do not make the prediction reduce to its inputs by construction, so the circularity score is 0.

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

The central claim is empirical and rests on a new dataset plus a new architecture; no mathematical derivation is attempted. The model components are assembled from Mamba, DiT, and standard losses, and no new physical entities are introduced.

free parameters (4)
  • Loss weight lambda = not specified
    Balances Ldiff and Lrecon in Eq. 10; hand-chosen, value omitted, affects reported metrics.
  • Timestep range Tdiff = not specified
    Uniform sampling U(1, Tdiff) for noise augmentation in Section 4.3; not reported.
  • Number of Mamba layers L and feature dimension D = not specified
    Architecture capacity for the temporal encoder in Section 4.2; hand-chosen, not reported.
  • DiT patch size and 12-block depth = patch size not specified, depth 12
    Diffusion module hyperparameters in Section 4.3; depth is stated as 12 but patch size is not.
assumptions (3)
  • domain assumption CHDL frame order represents temporal evolution
    Section 3.1 captures frames at 10 equidistant positions during stage translation; the paper treats this as a video prediction task.
  • domain assumption Denoising the input frames transfers detail to future target frames
    Section 4.3 trains epsilon_theta on noised input X and at inference adds the t=0 residual to Y_mamba; no evidence establishes this transfer.
  • domain assumption Bilinear interpolation of discarded abnormal frames preserves task-relevant dynamics
    Section 3.1 reconstructs approximately 1.3% of frames; metrics computed on interpolated data may overstate fidelity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DIFFUMA: High-Fidelity Spatio-Temporal Video Prediction via Dual-Path Mamba and Diffusion Enhancement." pith.science (2026). https://pith.science/paper/CRN7A6S3

@misc{pith2026250706738,
  author       = {Pith},
  title        = {Pith review of: DIFFUMA: High-Fidelity Spatio-Temporal Video Prediction via Dual-Path Mamba and Diffusion Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CRN7A6S3}},
  note         = {Machine review of arXiv:2507.06738}
}
read the original abstract

Spatio-temporal video prediction plays a pivotal role in critical domains, ranging from weather forecasting to industrial automation. However, in high-precision industrial scenarios such as semiconductor manufacturing, the absence of specialized benchmark datasets severely hampers research on modeling and predicting complex processes. To address this challenge, we make a twofold contribution.First, we construct and release the Chip Dicing Lane Dataset (CHDL), the first public temporal image dataset dedicated to the semiconductor wafer dicing process. Captured via an industrial-grade vision system, CHDL provides a much-needed and challenging benchmark for high-fidelity process modeling, defect detection, and digital twin development.Second, we propose DIFFUMA, an innovative dual-path prediction architecture specifically designed for such fine-grained dynamics. The model captures global long-range temporal context through a parallel Mamba module, while simultaneously leveraging a diffusion module, guided by temporal features, to restore and enhance fine-grained spatial details, effectively combating feature degradation. Experiments demonstrate that on our CHDL benchmark, DIFFUMA significantly outperforms existing methods, reducing the Mean Squared Error (MSE) by 39% and improving the Structural Similarity (SSIM) from 0.926 to a near-perfect 0.988. This superior performance also generalizes to natural phenomena datasets. Our work not only delivers a new state-of-the-art (SOTA) model but, more importantly, provides the community with an invaluable data resource to drive future research in industrial AI.

Figures

Figures reproduced from arXiv: 2507.06738 by the authors.

Figure 1
Figure 1. Qualitative and Quantitative Comparison of Pre￾diction Performance. arXiv:2507.06738v1 [cs.CV] 9 Jul 2025 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the CHDL Dataset Acquisition and Structure. (a) The data acquisition system, illustrating how the high-precision dicing machine (ADT-8230) is integrated with the camera in the automatic alignment system to capture microscopic images of the dicing lane during wafer translation. (b) An example of a CHDL data sample. Each sample consists of 10 consecutive frames, divided into an input sequence x (the first … view at source ↗
Figure 3
Figure 3. An overview of the DiffuMa. In this framework, the Mamba module models long-range temporal context by utilizing multi-layer [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Performance Comparison: A Comprehensive Analysis of MSE, MAE and SSIM Trends Across Epochs [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Predictive Performance of DiffuMa on WeatherBench [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Predictive Performance of DiffuMa on CHDL [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Comparative Analysis of Prediction Results on WeatherBench [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Comparative Analysis of Prediction Details on Weather [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Comparative Analysis of Prediction Details on CHDL [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 34 canonical work pages

  1. [1]

    H. Wu, F. Xu, C. Chen, X.-S. Hua, X. Luo, H. Wang, Pastnet: Introducing physical inductive biases for spatio-temporal video prediction, in: Proceedings of the 32nd ACM International Conference on Multi- media, 2024, pp. 2917–2926

    work page 2024

  2. [2]

    H. Wu, Y. Liang, W. Xiong, Z. Zhou, W. Huang, S. Wang, K. Wang, Earthfarsser: Versatile spatio- temporal dynamical systems modeling in one model, in: Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 2024, pp. 15906–15914. 12

    work page 2024

  3. [3]

    H. Li, W. Fu, H. Zhang, W. Liu, S. Sun, T. Zhang, Spatio–temporal graph hierarchical learning frame- work for metro passenger flow prediction across sta- tions and lines, Knowledge-Based Systems (2025) 113132

    work page 2025

  4. [4]

    W. Liu, M. J. Pyrcz, Physics-informed graph neural network for spatial-temporal production forecasting, Geoenergy Science and Engineering 223 (2023) 211486

    work page 2023

  5. [5]

    G. Li, Z. Zhao, X. Guo, L. Tang, H. Zhang, J. Wang, Towards integrated and fine-grained traffic forecasting: A spatio-temporal heterogeneous graph transformer approach, Information Fusion 102 (2024) 102063

    2024

  6. [6]

    Z. Lin, M. Li, Z. Zheng, Y. Cheng, C. Yuan, Self- attention convlstm for spatiotemporal prediction, in: Proceedings of the AAAI conference on artificial intel- ligence, volume 34, 2020, pp. 11531–11538

    work page 2020

  7. [7]

    Z. Wu, S. Pan, G. Long, J. Jiang, C. Zhang, Graph wavenet for deep spatial-temporal graph modeling, in: International Joint Conference on Artificial Intelli- gence, 2021, pp. 1907–1913

    work page 2021

  8. [8]

    J. Gao, X. Li, R. Jiang, X. Song, S. Shang, Y. Yao, Z. Wang, Spatio-temporal meta-learning for urban traffic prediction, IEEE Transactions on Knowledge and Data Engineering 35 (2023) 1215–1228

    work page 2023

Show all 48 references
  1. [9]

    L. Zhao, Y. Song, C. Zhang, Y. Liu, P. Wang, T. Lin, M. Deng, H. Li, T-gcn: A temporal graph convolu- tional network for traffic prediction, IEEE transac- tions on intelligent transportation systems 21 (2019) 3848–3858

    2019

  2. [10]

    A. A. Aleissaee, A. Kumar, R. M. Anwer, S. Khan, H. Cholakkal, G.-S. Xia, F. S. Khan, Transformers in remote sensing: A survey, Remote Sensing 15 (2023) 1860

    2023

  3. [11]

    S. Du, T. Yang, F. Teng, J. Zhang, T. Li, Y. Zheng, Multi-scale feature enhanced spatio-temporal learning for traffic flow forecasting, Knowledge-Based Systems 294 (2024) 111787

    2024

  4. [12]

    Hochreiter, J

    S. Hochreiter, J. Schmidhuber, Long short-term mem- ory, Neural Computation 9 (1997) 1735–1780

    1997

  5. [13]

    K. Cho, B. Van Merriënboer, C. Gulcehre, D. Bah- danau, F. Bougares, H. Schwenk, Y. Bengio, Learn- ing phrase representations using rnn encoder-decoder for statistical machine translation, arXiv preprint arXiv:1406.1078 (2014)

    2014 arXiv

  6. [14]

    Y. Seo, M. Defferrard, P. Vandergheynst, X. Bresson, Structured sequence modeling with graph convolu- tionalrecurrentnetworks, in: InternationalConference on Neural Information Processing, 2018, pp. 362–373

    2018

  7. [15]

    X. Shi, Z. Chen, H. Wang, D.-Y. Yeung, W.-K. Wong, W.-c. Woo, Convolutional lstm network: A machine learning approach for precipitation nowcasting, Ad- vances in neural information processing systems 28 (2015)

    2015

  8. [16]

    Y. Wang, L. Jiang, M.-H. Yang, L.-J. Li, M. Long, L. Fei-Fei, Eidetic 3d lstm: A model for video pre- diction and beyond, in: International conference on learning representations, 2018

    2018

  9. [17]

    Z. Wu, S. Pan, G. Long, J. Jiang, C. Zhang, Graph wavenet for deep spatial-temporal graph modeling, arXiv preprint arXiv:1906.00121 (2019)

    2019 arXiv

  10. [18]

    G. Jin, Y. Liang, Y. Fang, Z. Shao, J. Huang, J. Zhang, Y. Zheng, Spatio-temporal graph neural networks for predictive learning in urban computing: A survey, IEEE Transactions on Knowledge and Data Engineer- ing 36 (2023) 5388–5408

    2023

  11. [19]

    T. N. Kipf, M. Welling, Semi-supervised classification with graph convolutional networks, in: ICLR, 2017

    2017

  12. [20]

    B. Yu, H. Yin, Z. Zhu, Spatio-temporal graph con- volutional networks: A deep learning framework for traffic forecasting, arXiv preprint arXiv:1709.04875 (2017)

    2017 arXiv

  13. [21]

    Sankar, X

    A. Sankar, X. Zhang, A. Krishnan, J. Han, Graph neural processes for spatio-temporal extrapolation, in: International Conference on Learning Representations, 2020

    2020

  14. [22]

    B. Cao, C. Li, Y. Song, Y. Qin, C. Chen, Network in- trusion detection model based on cnn and gru, Applied Sciences 12 (2022) 4184

    2022

  15. [23]

    L. Bai, L. Yao, C. Li, X. Wang, C. Wang, Adaptive graph convolutional recurrent network for traffic fore- casting, Advances in neural information processing systems 33 (2020) 17804–17815

    2020

  16. [24]

    Z. L. Li, G. W. Zhang, J. Yu, L. Y. Xu, Dynamic graph structure learning for multivariate time series forecasting, Pattern Recognition 138 (2023) 109423

    2023

  17. [25]

    H. Wu, J. Xu, J. Wang, M. Long, Autoformer: De- composition transformers with Auto-Correlation for long-term series forecasting, in: NeurIPS, 2021

    2021

  18. [26]

    Z. Gao, X. Shi, H. Wang, Y. Zhu, Y. B. Wang, M. Li, D.-Y. Yeung, Earthformer: Exploring space-time transformers for earth system forecasting, Advances in Neural Information Processing Systems 35 (2022) 25390–25403

    2022

  19. [27]

    H. Wu, J. Wu, J. Xu, J. Wang, M. Long, Flow- former: Linearizing transformers with conservation flows, arXiv preprint arXiv:2202.06258 (2022). 13

    2022 arXiv

  20. [28]

    B. Lim, S. Ö. Arık, N. Loeff, T. Pfister, Temporal fu- sion transformers for interpretable multi-horizon time series forecasting, International Journal of Forecasting 37 (2021) 1748–1764

    2021

  21. [29]

    Shehzad, F

    A. Shehzad, F. Xia, S. Abid, C. Peng, S. Yu, D. Zhang, K. Verspoor, Graph transformers: A survey, arXiv preprint arXiv:2407.09777 (2024)

    2024

  22. [30]

    A. Gu, T. Dao, Mamba: Linear-time sequence mod- eling with selective state spaces, arXiv preprint arXiv:2312.00752 (2023)

    2023 arXiv

  23. [31]

    A. Gu, I. Johnson, K. Goel, K. Saab, T. Dao, A. Rudra, C. Ré, Combining recurrent, convolutional, and continuous-time models with linear state-space lay- ers, in: ICLR, 2022

    2022

  24. [32]

    J. T. Smith, A. Warrington, S. W. Linderman, Sim- plified state space layers for sequence modeling, arXiv preprint arXiv:2208.04933 (2022)

    2022 arXiv

  25. [33]

    Karras, M

    T. Karras, M. Aittala, T. Aila, S. Laine, Elucidating the design space of diffusion-based generative models, Advances in neural information processing systems 35 (2022) 26565–26577

    2022

  26. [34]

    Tashiro, J

    Y. Tashiro, J. Song, Y. Song, S. Ermon, Csdi: Con- ditional score-based diffusion models for probabilistic time series imputation, Advances in neural informa- tion processing systems 34 (2021) 24804–24816

    2021

  27. [35]

    Y. Yang, M. Jin, H. Wen, C. Zhang, Y. Liang, L. Ma, Y. Wang, C. Liu, B. Yang, Z. Xu, et al., A survey on diffusion models for time series and spatio-temporal data, arXiv preprint arXiv:2404.18886 (2024)

    2024

  28. [36]

    H. Wu, C. Wang, F. Xu, J. Xue, C. Chen, X.-S. Hua, X. Luo, Pure: Prompt evolution with graph ode for out-of-distribution fluid dynamics modeling, Advances in Neural Information Processing Systems 37 (2024) 104965–104994

    2024

  29. [37]

    Y. Wang, Z. Gao, M. Long, J. Wang, S. Y. Philip, Predrnn++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning, in: International conference on machine learning, PMLR, 2018, pp. 5123–5132

    2018

  30. [38]

    H. Wu, Y. Gao, R. Shu, K. Wang, R. Gou, C. Wu, X. Liu, J. He, S. Cao, J. Fang, X. Shi, F. Tao, Q. Song, S. Ji, Y. Xiang, Y. Sun, J. Li, F. Xu, H. Dong, H. Wang, F. Zhang, P. Zhao, X. Wu, Q. Wen, D. Chen, X. Huang, Advanced long-term earth system forecast- ing by learning the s...

    2025

  31. [39]

    H. Wu, Y. Gao, R. Shu, Z. Han, F. Xu, Z. Zhu, Q. Wen, X. Wu, K. Wang, X. Huang, Turb-l1: Achieving long-term turbulence tracing by tackling spectral bias, arXiv preprint arXiv:2505.19038 (2025)

    2025

  32. [40]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weis- senborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al., An im- age is worth 16x16 words: Transformers for image recognition at scale, arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  33. [41]

    K. He, X. Chen, S. Xie, Y. Li, P. Dollár, R. Girshick, Masked autoencoders are scalable vision learners, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 16000–16009

    2022

  34. [42]

    Sohl-Dickstein, E

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, S. Ganguli, Deep unsupervised learning using nonequi- librium thermodynamics, in: International conference on machine learning, pmlr, 2015, pp. 2256–2265

    2015

  35. [43]

    Srivastava, E

    N. Srivastava, E. Mansimov, R. Salakhudinov, Un- supervised learning of video representations using lstms, in: International conference on machine learn- ing, PMLR, 2015, pp. 843–852

    2015

  36. [44]

    Schuldt, I

    C. Schuldt, I. Laptev, B. Caputo, Recognizing human actions: a local svm approach, in: Proceedings of the 17th International Conference on Pattern Recognition,

  37. [45]

    Soomro, A

    K. Soomro, A. R. Zamir, M. Shah, Ucf101: A dataset of 101 human actions classes from videos in the wild, arXiv preprint arXiv:1212.0402 (2012)

    2012 arXiv

  38. [46]

    Cordts, M

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. En- zweiler, R. Benenson, U. Franke, S. Roth, B. Schiele, The cityscapes dataset for semantic urban scene un- derstanding, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3213–3223

    2016

  39. [47]

    S. Rasp, P. D. Dueben, S. Scher, J. A. Weyn, S. Mouatadid, N. Thuerey, Weatherbench: a bench- mark data set for data-driven weather forecasting, Journal of Advances in Modeling Earth Systems 12 (2020) e2020MS002203. 14

    2020

  40. [2004]

    ICPR 2004., volume 3, IEEE, 2004, pp. 32–36

    2004

Pith tools

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