Pith. sign in

REVIEW 4 major objections 3 minor 51 references

This paper argues that the missing lever in training-free diffusion feature caching is not a better forecaster but a way to decide, at runtime, when and how much to trust each forecast.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A disagreement signal between two cached forecasters lets diffusion sampling decide when to trust a forecast, improving quality at matched denoiser evaluations.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Genuinely new reliability signal for feature caching, but the calibration-trace overlap question has to be answered before the matched-NFE gains are fully trusted. the 4 major comments →

arxiv 2608.01740 v1 pith:V7IDQZ5N submitted 2026-08-03 cs.LG cs.AI

Disagree to Accelerate: Closing the Loop on Diffusion Feature Forecasts

Yanchao Li , Jiaqing Xie , Ben Gao , Wanhao Liu , Yanbo Wang , T. Y. Tsui , Jinfei Liu , Yuqiang Li
show 1 more author
Tianfan Fu
This is my paper
classification cs.LG cs.AI
keywords diffusion sampling accelerationfeature cachingtraining-free forecastingclosed-loop controlforecast reliabilitydisagreement signaldenoiser evaluation budgetRACER
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 claims that forecast reliability can be read from the cache itself: two cheap forecasts with different biases agree on smooth stretches of the feature trajectory and diverge where prediction turns hard, and their normalized disagreement flags high-error steps without any extra denoiser evaluation. On this signal, RACER closes the loop with two responses: it continuously shrinks uncertain forecasts toward the last computed feature, and at the riskiest steps it spends one real evaluation and repays it by skipping a later scheduled one, so the per-prompt evaluation count stays exactly fixed. At matched NFE, RACER improves the strongest open-loop baseline on SD3.5-Large, FLUX.1-dev, Wan2.1-14B, and HunyuanVideo across DrawBench, VBench, and COCO, and on SD3.5 it reaches equal quality at lower latency. The central claim is that reliable diffusion acceleration depends on how forecasts are used, not only on how they are made.

Core claim

The central discovery is that the reliability of a diffusion feature forecast is observable from the cache alone. The paper defines a runtime signal r_t = ||h_hat_t - g_hat_t|| / ||h_hat_t|| between a base forecaster and a second forecaster with a different bias, called the observer. Because the two forecasts carry different biases, their disagreement grows exactly where the feature trajectory bends and the true forecast error spikes. Across four image and video models, this signal identifies the top-20% forecast-error steps at a mean AUROC of 0.94, versus 0.77 for the input-side signals that prior caches use. RACER converts this signal into a closed-loop controller: an exponential-sigmoid t

What carries the argument

The load-bearing objects are the disagreement signal r_t and the trust function kappa_t. r_t is the relative norm difference between two forecasts produced from the same cache—the base forecaster and an observer with a different bias—so reliability becomes visible with O(F) arithmetic and no extra denoiser evaluation. kappa_t = exp(-lambda * max(k_t - 1, 0)) * sigmoid(beta * (theta_kappa - r_t)) sets the interpolation weight between the anchor feature and the forecast; kappa = 0 recovers pure feature reuse and kappa = 1 recovers full open-loop forecasting. A refresh-and-repay debt counter turns flagged risky steps into an exact budget-conservation mechanism, keeping the number of denoiser ev

Load-bearing premise

The disagreement signal is a faithful reliability proxy only if the observer is a valid check: its own error must be bounded and only weakly correlated with the base forecaster's error, otherwise the two forecasts can agree while both are badly wrong.

What would settle it

Compute the true base forecast error and observer error step-by-step on many prompts and measure their correlation rho_fo and the observer error bound B_o. If rho_fo is high on hard steps (say above 0.8) or B_o is comparable to the base error, then Equations (4)-(5) show disagreement cannot reveal forecast error, so an ablation replacing r_t with the true error should show RACER's gains shrinking or reversing.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If RACER's central claim is correct, training-free feature caching can be improved without modifying the denoiser and without training: adding a second forecaster and a controller is enough.
  • The gains concentrate where open-loop caches break—aggressive acceleration regimes—so the same fixed NFE budget buys more quality exactly when skipping is most risky.
  • The controller is decoupled from the base forecaster: it improves a Taylor base as well as a Chebyshev base, so the benefit transfers across forecasting designs.
  • The per-prompt NFE remains exactly the base schedule's NFE, meaning the improvement is not additional compute but a reallocation of the same compute to the most unreliable steps.
  • The signal and controller transfer across models and datasets, with image-side calibration transferring from SD3.5 to FLUX with only a small trust-threshold change.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if disagreement reliably flags hard steps, the same signal could also choose where to place the computed steps U itself, rather than only deciding whether to refresh within a fixed schedule.
  • Editorial inference: the validity conditions suggest a practical diagnostic—measuring the observer's error bound and its error correlation with the base on a short trace would predict when the disagreement signal will fail.
  • Editorial inference: because the signal only needs two forecasts with different biases, one could dynamically select among multiple base forecasters or observers rather than fixing a single pair.
  • Editorial inference: the query-by-committee motivation hints at an ensemble interpretation, but the paper stops at using disagreement as a risk score and does not develop ensemble weighting of multiple forecasts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. RACER is a training-free closed-loop controller for diffusion feature caching. The paper observes that two cached forecasts with different biases (a Chebyshev base and a Taylor observer) agree on smooth stretches and diverge at hard steps; this disagreement r_t is proposed as a reliability signal requiring no extra denoiser evaluations. RACER maps r_t to two responses: continuous shrinkage of the trusted feature toward the anchor via κ_t = exp(-λ max(k_t-1,0)) σ(β(θ_κ - r_t)), and an exact refresh-and-repay mechanism that recomputes a feature at high-disagreement steps and skips a later scheduled step to keep per-prompt NFE constant. The paper proves three propositions: a triangle-inequality error bound, an MSE-optimal shrinkage formula, and exact budget conservation. Experiments compare RACER against open-loop baselines on SD3.5-Large, FLUX.1-dev, Wan2.1-14B, and HunyuanVideo on DrawBench, VBench, and COCO, reporting matched-NFE gains, ablations, observer choice, and transfer to a Taylor base.

Significance. If the empirical results hold, the contribution is a simple, orthogonal control layer that can be stacked on any feature forecaster. The disagreement signal is cheap, the refresh-and-repay budget argument is clean, and the paper includes multiple ablations and a public code link. The theoretical results are elementary (Prop. 1 is a convexity/triangle inequality, Prop. 2 is standard forecast combination, Prop. 3 is structural), so the value rests on the empirical validation. The benchmark sweep across four models and three datasets is welcome, and the transfer experiments to a Taylor base help establish generality.

major comments (4)
  1. [Experimental Setup, 'Implementation details'; Method, 'Closed-Loop Control'] The matched-NFE gains in Tables 1–3 and Figure 3 are in-sample unless the offline trace used to fix λ, β, θ_κ, θ_ref is disjoint from the evaluation prompts. The paper says 'A short offline trace fixes the trust scalars' and 'video thresholds are calibrated per model and regime on a short trace' but never states the trace composition or overlap status. This is load-bearing because the central claim is a comparison at matched NFE over specific benchmarks; quantile calibration and threshold selection on the evaluation distribution would directly inflate the reported improvements. Please specify the trace (number of prompts, source, disjointness) for all models, and for image-side transfer from SD3.5 to FLUX.
  2. [Theoretical Analysis, Proposition 1] The claim that this is a 'deterministic error bound' is stronger than what is established. B_f and B_h are not defined in the main text; on real traces they are necessarily measured per trace. If B_f and B_h are empirical maxima from a calibration trace, the inequality holds for those traces by construction and gives no distribution-free guarantee for new prompts. The paper should either define B_f and B_h explicitly and provide a priori bounds (e.g., Lipschitz or Chebyshev remainder bounds) or reframe the result as an empirical envelope. This affects the abstract's 'deterministic error bound' claim.
  3. [Theoretical Analysis, Eqs. (4)–(5)] The theoretical validity of the disagreement signal requires an observer error bound B_o and a weak error correlation ρ_fo. The paper never reports B_o for the deployed Taylor observers, and reports error correlations only on 'worst steps' (0.17–0.44). Without B_o, Eq. (4) cannot certify that large disagreement implies large base error. The empirical AUROC of 0.94 is supportive but does not substitute for the stated theoretical conditions. Please report B_o estimates on the calibration trace or explicitly present (4)–(5) as motivating assumptions rather than verified bounds.
  4. [Table 1, α=5.0 and α=7.0 rows] The FORA, TaylorSeer, TeaCache, and ToCa entries are identical to the α=3.0 rows despite different reported NFE (9 and 8 vs. ~10). The note 'The four cache baselines land on one schedule for α=3.0 and 5.0' does not explain how the same PSNR/SSIM can accompany different NFE. If these baseline numbers were not re-measured at the aggressive schedules, the matched-NFE comparison at α=5.0/7.0 is not supported and the claim 'advantage holds in every regime' is compromised. Please correct the table or clarify the scheduling.
minor comments (3)
  1. [Throughout] There are small spacing typos such as '2.6to5.4×' and '0.76×the latency'; these should be fixed. Also, Eq. (2) should explicitly state that the denominator is the norm of the base forecast, not the true feature.
  2. [Method, Preliminaries] The Taylor observer g is not defined in the main text. Please give the exact order and step-size formula used for the Taylor expansion, or cite where it is specified.
  3. [Experimental Setup] The phrase 'image parameters are transferred from SD3.5 to FLUX' is vague: it should state which parameters are transferred, and whether the transfer includes the refresh threshold or only the trust scalars.

Circularity Check

0 steps flagged

No demonstrated circularity: the disagreement signal is empirically validated, the propositions are independent bounds or standard combination results, and the only flagged issue is that the calibration trace is not stated to be disjoint from evaluation prompts (a transparency risk, not a circular reduction).

full rationale

The central mechanism is not circular. The disagreement signal r = ||h_hat - g_hat||/||h_hat|| (Eq. 2) is an empirical proxy whose validity is tested externally: it flags top-20% error steps at AUROC 0.94 versus 0.77 for input-side signals (Figure 1). This is a falsifiable measurement, not an equivalence by definition. Proposition 1 is a triangle-inequality/convex-interpolation bound: because y_hat(kappa) = a + kappa(h_hat - a), the bound ||y_hat(kappa)-h|| <= kappa Bf + (1-kappa)Bh is a direct consequence and is used only as a bound, with tightness measured on traces; it is not used to derive the claimed quality gains. Proposition 2 is a standard forecast-combination result (cited to Wang et al. 2023), and the actual controller in Eq. (3) is a calibrated exponential-sigmoid surrogate, not claimed to equal kappa*; the paper explicitly calls horizon and r 'proxies.' Proposition 3 is an exact-budget guarantee that holds by construction of the debt counter ('the count holds by construction'), so it is a design invariant rather than a prediction being validated. No load-bearing self-citation is present: the Chebyshev base is an external baseline (Han et al. 2026), and the theoretical results cite external literature (Seung et al.; Wang et al.). The one legitimate concern is that the paper says 'A short offline trace fixes the trust scalars' and 'video thresholds are calibrated per model and regime on a short trace' without stating that this trace is disjoint from the DrawBench/VBench/COCO evaluation prompts. If the trace overlaps the evaluation distribution, the matched-NFE gains in Tables 1-3 could be in-sample. However, the paper also does not state overlap, and the signal's AUROC and the component ablations (Tables 4-5) provide independent support for the mechanism. This is an unverified-protocol / correctness risk, not an exhibited reduction of the results to the fitted scalars, so it does not constitute demonstrated circularity. Score 1 reflects a largely self-contained derivation with one unspecified calibration-trace protocol.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The paper introduces no new physical or conceptual entities beyond the RACER controller itself. The load-bearing assumptions are the observer validity conditions in Eqs. (4) and (5), the empirical validity of the disagreement proxy, and the hand-fitted controller scalars.

free parameters (5)
  • lambda (horizon decay rate) = not reported
    Set by short offline trace; controls how fast trust shrinks with distance from the last computed step (Eq. 3).
  • beta (sigmoid steepness) = not reported
    Set by short offline trace; controls the sharpness of the trust transition around the threshold (Eq. 3).
  • theta_kappa (trust threshold) = not reported
    Shift of the sigmoid in Eq. 3; image-side values transfer from SD3.5 to FLUX with a reported 4.7% change.
  • theta_ref (refresh threshold) = not reported
    Above this disagreement the controller spends a real denoiser evaluation to refresh the feature.
  • Bf and Bh (forecast and anchor error bounds) = empirically estimated on traces
    Used in Proposition 1 to state the deterministic error bound; the values are measured per trace, not derived a priori.
axioms (6)
  • standard math Triangle inequality for the interpolation error (Proposition 1)
    Used to state the deterministic error bound; true for any norm.
  • standard math Finite error second moments and standard MSE-optimal forecast combination formula (Wang et al. 2023)
    Proposition 2 relies on the standard correlated-estimate combination result.
  • domain assumption A large disagreement forces a large base forecast error only when the observer error is bounded by B_o (Eq. 4)
    The reliability signal is only as good as this bound; no actual B_o is supplied for the deployed observers.
  • domain assumption A large base forecast error surfaces in the disagreement only when forecast and observer errors have low correlation (Eq. 5)
    The paper reports worst-step correlations of 0.17 to 0.44, but the validity claim requires this to hold generally.
  • domain assumption Two forecasters with different biases (Chebyshev base, Taylor observer) provide a valid disagreement signal
    Core observational claim; supported empirically by AUROC but not proven.
  • ad hoc to paper Fixed regime rule: use step coordinate for alpha<=0.75 and logSNR coordinate for alpha>=3.0
    Hand-chosen coordinate switch with no derivation for the cutoff; affects all aggressive regimes.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Disagree to Accelerate: Closing the Loop on Diffusion Feature Forecasts." pith.science (2026). https://pith.science/paper/V7IDQZ5N

@misc{pith2026260801740,
  author       = {Pith},
  title        = {Pith review of: Disagree to Accelerate: Closing the Loop on Diffusion Feature Forecasts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V7IDQZ5N}},
  note         = {Machine review of arXiv:2608.01740}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Training-free feature forecasting accelerates diffusion sampling by predicting features at skipped denoising steps. Recent work has mainly focused on designing stronger forecasters. Yet forecast error varies sharply across steps, and open-loop caches trust the forecast in full at every skipped step. This fixed trust is what breaks as acceleration turns aggressive. The missing question is not only how to forecast better, but when and how much to trust a forecast. We show that reliability can be observed from the cache itself. Two forecasts agree where the feature trajectory is smooth, and they diverge where prediction turns hard. Their disagreement is a cheap runtime signal, and it costs no extra denoiser evaluation. Based on this signal, we introduce RACER, a training-free closed-loop controller with two responses. It continuously shrinks uncertain forecasts toward the last computed feature. At the riskiest steps, RACER refreshes the feature and repays the added evaluation by skipping a later scheduled one. We derive a deterministic error bound for the shrinkage and empirically evaluate its validity and tightness across acceleration regimes. At the same number of denoiser evaluations, RACER improves the strongest open-loop baseline across SD3.5-Large, FLUX.1-dev, Wan2.1-14B, and HunyuanVideo on DrawBench, VBench, and COCO. On SD3.5, we further show that RACER samples faster at equal quality. RACER generalizes across forecasting designs as well. For example, it recovers much of the quality lost on a Taylor base. These results show that reliable diffusion acceleration also depends on how forecasts are used. Code is available at https://github.com/LiZaiyuan0619/RACER

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 1 canonical work pages

  1. [1]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Han, Jiaqi and Shi, Juntong and Li, Puheng and Ye, Haotian and Guo, Qiushan and Ermon, Stefano , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2026 , pages =

    2026

  2. [2]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Liu, Jiacheng and Zou, Chang and Lyu, Yuanhuiyi and Chen, Junjie and Zhang, Linfeng , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2025 , pages =

    2025

  3. [3]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Ma, Xinyin and Fang, Gongfan and Wang, Xinchao , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  4. [4]

    2024 , eprint=

    FORA: Fast-Forward Caching in Diffusion Transformer Acceleration , author=. 2024 , eprint=

    2024

  5. [5]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages=

    Compute only 16 tokens in one timestep: Accelerating diffusion transformers with cluster-driven feature caching , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=

  6. [6]

    The Thirteenth International Conference on Learning Representations , year=

    Accelerating Diffusion Transformers with Token-wise Feature Caching , author=. The Thirteenth International Conference on Learning Representations , year=

  7. [7]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Liu, Feng and Zhang, Shiwei and Wang, Xiaofeng and Wei, Yujie and Qiu, Haonan and Zhao, Yuzhong and Zhang, Yingya and Ye, Qixiang and Wan, Fang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2025 , pages =

    2025

  8. [8]

    2026 , eprint=

    SoftCap: Soft-Budget Control for Diffusion Transformer Acceleration , author=. 2026 , eprint=

    2026

  9. [9]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages =

    Liu, Jiacheng and Zou, Chang and Lyu, Yuanhuiyi and Ren, Fei and Wang, Shaobo and Li, Kaixin and Zhang, Linfeng , title =. Proceedings of the 33rd ACM International Conference on Multimedia , pages =. 2025 , isbn =. doi:10.1145/3746027.3755331 , abstract =

    work page arXiv 2025

  10. [10]

    The Fourteenth International Conference on Learning Representations , year=

    HiCache: A Plug-in Scaled-Hermite Upgrade for Taylor-Style Cache-then-Forecast Diffusion Acceleration , author=. The Fourteenth International Conference on Learning Representations , year=

  11. [11]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Forecast then calibrate: Feature caching as ode for efficient diffusion transformers , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  12. [12]

    2025 , eprint=

    Less is Enough: Training-Free Video Diffusion Acceleration via Runtime-Adaptive Caching , author=. 2025 , eprint=

    2025

  13. [13]

    Forty-third International Conference on Machine Learning , year=

    Budget-Constrained Step-Level Diffusion Caching , author=. Forty-third International Conference on Machine Learning , year=

  14. [14]

    2026 , eprint=

    SeaCache: Spectral-Evolution-Aware Cache for Accelerating Diffusion Models , author=. 2026 , eprint=

    2026

  15. [15]

    The Fourteenth International Conference on Learning Representations , year=

    DiCache: Let Diffusion Model Determine Its Own Cache , author=. The Fourteenth International Conference on Learning Representations , year=

  16. [16]

    International Conference on Learning Representations , year=

    Denoising Diffusion Implicit Models , author=. International Conference on Learning Representations , year=

  17. [17]

    2022 , url=

    Cheng Lu and Yuhao Zhou and Fan Bao and Jianfei Chen and Chongxuan Li and Jun Zhu , booktitle=. 2022 , url=

    2022

  18. [18]

    Wenliang Zhao and Lujia Bai and Yongming Rao and Jie Zhou and Jiwen Lu , booktitle=. Uni. 2023 , url=

    2023

  19. [19]

    2023 , eprint=

    Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference , author=. 2023 , eprint=

    2023

  20. [20]

    2022 , eprint=

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author=. 2022 , eprint=

    2022

  21. [21]

    2023 , eprint=

    Token Merging for Fast Stable Diffusion , author=. 2023 , eprint=

    2023

  22. [22]

    2024 , eprint=

    DistriFusion: Distributed Parallel Inference for High-Resolution Diffusion Models , author=. 2024 , eprint=

    2024

  23. [23]

    2024 , eprint=

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis , author=. 2024 , eprint=

    2024

  24. [24]

    2025 , eprint=

    Wan: Open and Advanced Large-Scale Video Generative Models , author=. 2025 , eprint=

    2025

  25. [25]

    2025 , eprint=

    HunyuanVideo: A Systematic Framework For Large Video Generative Models , author=. 2025 , eprint=

    2025

  26. [26]

    Advances in neural information processing systems , volume=

    Photorealistic text-to-image diffusion models with deep language understanding , author=. Advances in neural information processing systems , volume=

  27. [27]

    2023 , eprint=

    VBench: Comprehensive Benchmark Suite for Video Generative Models , author=. 2023 , eprint=

    2023

  28. [28]

    2015 , eprint=

    Microsoft COCO: Common Objects in Context , author=. 2015 , eprint=

    2015

  29. [29]

    2018 , eprint=

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric , author=. 2018 , eprint=

    2018

  30. [30]

    2023 , eprint=

    ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation , author=. 2023 , eprint=

    2023

  31. [31]

    2021 , eprint=

    Learning Transferable Visual Models From Natural Language Supervision , author=. 2021 , eprint=

    2021

  32. [32]

    2026 , eprint=

    ReCache: Learning Budget-Aware Caching Schedules for Diffusion Models via REINFORCE , author=. 2026 , eprint=

    2026

  33. [33]

    2026 , eprint=

    ProCache: Constraint-Aware Feature Caching with Selective Computation for Diffusion Transformer Acceleration , author=. 2026 , eprint=

    2026

  34. [34]

    2026 , eprint=

    Predict to Skip: Linear Multistep Feature Forecasting for Efficient Diffusion Transformers , author=. 2026 , eprint=

    2026

  35. [35]

    Yu, Zichao and Zou, Zhen and Shao, Guojiang and Zhang, Chengwei and Xu, Shengze and Huang, Jie and Zhao, Feng and Cun, Xiaodong and Zhang, Wenyi , journal=

  36. [36]

    2025 , eprint=

    FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space , author=. 2025 , eprint=

    2025

  37. [37]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Li, Xiuyu and Liu, Yijiang and Lian, Long and Yang, Huanrui and Dong, Zhen and Kang, Daniel and Zhang, Shanghang and Keutzer, Kurt , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2023 , pages =

    2023

  38. [38]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Structural Pruning for Diffusion Models , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  39. [39]

    International Conference on Learning Representations , year=

    Progressive Distillation for Fast Sampling of Diffusion Models , author=. International Conference on Learning Representations , year=

  40. [40]

    Journal of the operational research society , volume=

    The combination of forecasts , author=. Journal of the operational research society , volume=. 1969 , publisher=

    1969

  41. [41]

    International Journal of Forecasting , volume=

    Forecast combinations: An over 50-year review , author=. International Journal of Forecasting , volume=. 2023 , publisher=

    2023

  42. [42]

    IEEE Transactions on Signal Processing , volume=

    KalmanNet: Neural network aided Kalman filtering for partially known dynamics , author=. IEEE Transactions on Signal Processing , volume=. 2022 , publisher=

    2022

  43. [43]

    1949 , publisher=

    Extrapolation, interpolation, and smoothing of stationary time series: with engineering applications , author=. 1949 , publisher=

    1949

  44. [44]

    2019 , publisher=

    Approximation theory and approximation practice, extended edition , author=. 2019 , publisher=

    2019

  45. [45]

    Proceedings of the fifth annual workshop on Computational learning theory , pages=

    Query by committee , author=. Proceedings of the fifth annual workshop on Computational learning theory , pages=

  46. [46]

    Advances in neural information processing systems , volume=

    Simple and scalable predictive uncertainty estimation using deep ensembles , author=. Advances in neural information processing systems , volume=

  47. [47]

    Denoising Diffusion Probabilistic Models , url =

    Ho, Jonathan and Jain, Ajay and Abbeel, Pieter , booktitle =. Denoising Diffusion Probabilistic Models , url =

  48. [48]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  49. [49]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  50. [50]

    IEEE transactions on image processing , volume=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE transactions on image processing , volume=. 2004 , publisher=

    2004

  51. [51]

    The Eleventh International Conference on Learning Representations , year=

    Flow Matching for Generative Modeling , author=. The Eleventh International Conference on Learning Representations , year=

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.