Pith. sign in

REVIEW 6 major objections 6 minor 70 references

Local Epistemic Uncertainty Guided Active Sampling for Plug-and-play Diffusive Image Restoration

T0 review · 6 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a pixel-wise epistemic uncertainty signal—the posterior covariance diagonal—can modulate prior strength and skip redundant steps in diffusion restoration, improving quality and cutting cost across baselines.

desk verdict A plausible plug-and-play trick that overclaims its theory and hides an infeasible computation; with major revisions it could be salvageable. read the letter →

arxiv 2608.06981 v1 pith:P6L6CI2F submitted 2026-08-07 cs.CV

Jiaqi Zhang , Zheng Pang , Rongrong Gao , Qiyuan Zhang , Yang Yang This is my paper · ORCID
classification cs.CV
keywords diffusionmodelimagerestorationplug-and-playlocalepistemicuncertaintyadaptivestepsizenull-spaceprojectiontrajectorypruningdataconsistency
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

Diffusion-based image restoration solvers usually push every pixel toward the same data-consistency constraint with the same step size at every time step. This paper argues that the reverse process itself reveals how confident the model is about each location: the diagonal of the posterior covariance of the clean-image estimate, derived from the Hessian of the model's log-probability. LEADer uses that per-pixel uncertainty to weaken or strengthen the prior locally, and to measure when the trajectory is stable enough to skip steps. If correct, the result is a plug-and-play wrapper that improves restoration quality of multiple baselines while cutting sampling time and adding almost no memory.

What carries the argument

Local epistemic uncertainty quantification is the central mechanism: the diagonal of $\boldsymbol{\Sigma}_t = \frac{1-\bar{\alpha}_t}{\bar{\alpha}_t}[\mathbf{I}-(1-\bar{\alpha}_t)\mathbf{F}_t]$, with $\mathbf{F}_t = -\nabla^2_{\mathbf{x}_t}\log p_t(\mathbf{x}_t)$ the model's Fisher information matrix, i.e., the second derivative of its log-probability at the current noisy state. Its per-pixel entries set the modulation strength in the Uncertainty-Calibrated Prior Modulation (UCPM) update, and its trace sets the admissible DDIM step size in State-Aware Trajectory Pruning (SATP) through the budget constraint $\delta(\Delta t) \approx (\Delta t)^\rho U_t \le B$.

What would settle it

Sample many noise predictions from the diffusion model at a fixed noisy state $\mathbf{x}_t$ and compare the empirical spread of the resulting $\mathbf{x}_{0|t}$ estimates, pixel by pixel, with the diagonal of $\boldsymbol{\Sigma}_t$ from Eq. (10); if the empirical covariance does not match, the uncertainty signal that drives both UCPM and SATP is not what the paper claims it is.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the local posterior covariance $\boldsymbol{\Sigma}_t = \frac{1-\bar{\alpha}_t}{\bar{\alpha}_t}[\mathbf{I}-(1-\bar{\alpha}_t)\mathbf{F}_t]$, where $\mathbf{F}_t = -\nabla^2_{\mathbf{x}_t}\log p_t(\mathbf{x}_t)$, is a usable control signal for both space and time in the reverse diffusion sampler. Spatially, a MAP formulation with this covariance as the precision scale yields an adaptive modulation matrix $\boldsymbol{\Lambda}_t^* = \boldsymbol{\Sigma}_t(\mathbf{I}+\mathbf{H}_\Phi \boldsymbol{\Sigma}_t)^{-1}$ that automatically attenuates prior intervention where the model is already confident; temporally, the trace $U_t = \frac{1}{N}\mathrm{Tr}(\boldsymbol{\Sigma}_t)$ bounds the local truncation error of a step, so the maximum safe step $\Delta t_{\mathrm{Active}}$ can be read off from $U_t$ under a budget $B$. Combined with null-space projection, the paper proves strict data consistency in the noise-free case and, under a Lipschitz drift, a deterministic global error bound for the pruned trajectory. The claim is that these two mechanisms are why the plug-and-play wrapper improves PSNR, LPIPS, and FID across the tested baselines while reducing function evaluations and wall-clock time.

Load-bearing premise

The load-bearing premise is that a pixel-level uncertainty map can be obtained in closed form from the second derivative of the diffusion model's log-probability at each state, and that the trace of that map bounds the error of skipping a sampling step; the paper states this without derivation and without specifying how such a large Hessian is computed for real images.

Editorial extensions

If this is right

  • Any diffusion restoration solver that produces a clean-image estimate $\mathbf{x}_{0|t}$ and uses null-space projection can be wrapped by LEADer without retraining, so the reported quality and speed gains should transfer to other null-space baselines beyond the six tested.
  • Because the modulation is confined to the null space, the data-consistency equation $\|\mathbf{y}-\mathbf{A}\hat{\mathbf{x}}\|_2^2 = 0$ holds exactly in the noise-free case, so per-pixel prior changes cannot corrupt the range-space measurements.
  • The deterministic error bound $\|\mathbf{x}_{\mathrm{Dense}}-\mathbf{x}_{\mathrm{LEADer}}\|_2 \le C\cdot B(e^{LT}-1)/L$ means the accelerated trajectory cannot drift arbitrarily far from the dense ODE solution, making the speedup compatible with a convergence guarantee.
  • The uncertainty trace drops and stabilizes as sampling proceeds, so SATP allocates more steps to early unstable phases and fewer to late stable ones; this is a distinct behavior from simply reducing the number of uniform steps.
  • The experiments show time savings of roughly 3–8% together with PSNR gains of 0.9–2.4% and double-digit LPIPS improvements, so the paper claims efficiency and quality together rather than trading one off against the other.

Reading between the lines

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

  • Editorial inference: the trace-based step selection is effectively an adaptive ODE controller; replacing the fixed budget $B$ with a learned or per-region budget could extend LEADer to latent diffusion models where the sampling manifold is lower-dimensional.
  • Editorial inference: the paper reports negligible memory overhead despite the $N\times N$ Hessian in Eq. (8), which implies the implementation approximates or avoids materializing the full Hessian; identifying that approximation is a natural next test, since Eq. (10) as written is not obviously computable at 256×256 resolution.
  • Editorial inference: the same uncertainty trace that drives step skipping could serve as a stopping criterion for early termination, cutting cost further; the paper's conclusion gestures at generative tasks, but a restoration-specific stopping rule is a direct extension.
  • Editorial inference: because UCPM reduces to no modulation where $\boldsymbol{\Sigma}_t \to 0$, it behaves like a self-sparsifying regularizer; this suggests a connection to adaptive sparsity and could make the method robust in mildly degraded regions where uniform priors tend to oversmooth.
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

6 major / 6 minor

Summary. The paper proposes LEADer, a plug-and-play module for diffusion-model-based image restoration that uses a local epistemic uncertainty estimate to adaptively modulate null-space prior updates and to prune sampling steps. The spatial component (UCPM) defines a Fisher-information-matrix-based covariance, uses it in a MAP objective with an unspecified prior, and projects the result into the null space. The temporal component (SATP) derives an adaptive step size from the trace of that covariance and claims a deterministic global error bound. Experiments on five restoration tasks over CelebA-HQ 1K and ImageNet 1K show consistent PSNR/LPIPS improvements and modest runtime reductions across several baselines, with very low reported memory overhead. The theoretical claims are not established as written: the covariance formula is asserted without derivation, the algorithm requires intractable N-by-N Hessians and inverses, the prior is never instantiated, and the error-bound proof is conditional on its own design goal.

Significance. The idea of using pixel-wise uncertainty to modulate null-space guidance and to adapt step sizes is useful and, if properly specified, could be a practical plug-and-play contribution to DMIR. The empirical evaluation is broad and internally consistent: it covers multiple degradation types, two datasets, several strong baselines, ablations, hyperparameter sensitivity, and memory measurements. However, the advertised theoretical guarantees are not independent supports for the method: strict data consistency is true by construction, and the deterministic error bound rests on an unproven relationship between the uncertainty trace and the actual local truncation error. The decisive weakness is reproducibility: the algorithm as written cannot run at the reported scale without an approximation that the paper never discloses, so the implementation used in the experiments is not defined by the text. If the authors can specify the approximation, correct or re-frame the theory accordingly, and provide full implementation details, the contribution would be considerably stronger.

major comments (6)
  1. [§4.2, Eqs. (8)–(10)] Equation (10) asserts that the posterior covariance is Σ_t = (1−ᾱ_t)/ᾱ_t [I − (1−ᾱ_t)F_t(x_t)], but no derivation is given, and this formula is not a consequence of Tweedie's formula or of a standard Laplace approximation of the marginal density p_t(x_t). Since F_t = −∇²_{x_t} log p_t(x_t) is the Hessian of the marginal density, it is not generally available from a score network and is not generally diagonal, although the text calls Σ_t a diagonal matrix. This relationship is the foundation of both UCPM and SATP, so the central theoretical claim is unsupported as written.
  2. [§4.3, Eq. (15); Algorithm 1 lines 8–9; Table 4] The UCPM update requires forming the N×N Hessian F_t and computing the inverse (I + H_Φ(x_{0|t})Σ_t)^{-1}. For 256×256 images this is a 65536×65536 dense object; storing Σ_t alone would cost roughly 16 GB in float32, and the matrix inverse is computationally infeasible. Table 4 nevertheless reports a memory increase of only 0.07–0.28%, so the implementation must rely on an approximation (diagonal, low-rank, trace-estimator, or similar) that is never described. Until that approximation is specified, Eq. (14) is not a well-defined algorithm and the theoretical MAP derivation does not apply to the implemented method.
  3. [§4.3, Eq. (11)] The regularizer Φ(x) in the MAP objective is never defined or instantiated in the experiments. The text gives only 'e.g., gradient and structural priors,' but neither ∇Φ(x_{0|t}) nor H_Φ(x_{0|t}) is specified for any experiment, so the closed-form update in Eq. (14) cannot be reproduced and the objective J(x) is not a well-defined function. This is a load-bearing gap because the whole UCPM modulation depends on the choice of this prior.
  4. [§4.4, Eq. (17)] The proportionality ∥d²x_t/dt²∥ ∝ Tr(Σ_t) is asserted without proof or reference. This is load-bearing because it is the only connection between the covariance formula in Eq. (10) and the adaptive step-size rule in Eq. (19). Without a proof of this relation, SATP is an ungrounded heuristic rather than a mechanism with the claimed error control.
  5. [§4.5, Proposition 4.2] The proof of Proposition 4.2 assumes that SATP bounds the single-step truncation error by C·B·Δt_k. That is exactly the property that Eq. (18) is supposed to establish, but Eq. (18) merely postulates δ(Δt) ≈ Δt^ρ · U_t ≤ B, and no proof connects U_t to the actual local truncation error of the reverse ODE. The Grönwall argument is standard, so the claimed deterministic global bound is conditional on the very assertion that needs to be proved.
  6. [§4.5, Proposition 4.1] Proposition 4.1 is true but vacuous: every estimate of the form A†y + P_N x_prior satisfies A x = y exactly because A P_N = 0. The result therefore holds by construction of the null-space projection in Eq. (20) and provides no independent guarantee or validation of the uncertainty modulation. Presenting it as a theoretical support for UCPM overstates what the property establishes.
minor comments (6)
  1. [Section 1 / Abstract] The code link appears as 'available at here' in Section 1 while the abstract gives a URL; please make the link explicit and consistent.
  2. [§4.4, Eq. (19)] The values of the scaling factor η and the numerical-stability constant ε are not reported; the sensitivity study in Table 3 sweeps only B, so the default configuration is not fully specified.
  3. [§4.4, Eq. (17)] Equation (17) contains two proportionalities; the sign of the term involving ∇² log p_t should be reconciled with Eq. (10) (I − (1−ᾱ_t)F_t versus I + (1−ᾱ_t)∇² log p_t) to avoid confusion.
  4. [Table 2] Table 2 contains a typo in the CS 25% row ('27.14.' with a stray decimal point).
  5. [Table 3] Table 3 shows identical PSNR and LPIPS values for B=0 and B=0.001; if this is not a typo, the text should explain why the smallest nonzero budget has no observable effect.
  6. [§B.2, Proof of Proposition 4.2] The proof assumes e_0 = 0 at the initial state, but Algorithm 1 starts from x_T ~ N(0,I), which is not the exact ODE solution; the discrepancy between the proof's initialization and the algorithm's stochastic initialization should be addressed.

Circularity Check

2 steps flagged · score 6.0 of 10

Two advertised theoretical guarantees reduce to definition and assumption: Proposition 4.1's strict data consistency is built into the null-space update by construction, and Proposition 4.2's deterministic error bound assumes SATP's own step-size constraint as its premise. The empirical comparisons are not circular.

  1. self definitional [Section 4.5, Proposition 4.1; Eq. (20) and Appendix B.1]
    "Since the modulated estimation of UCPM is defined as ˆx= A†y+𝑃𝑁 x𝑃𝑟𝑖𝑜𝑟, under the noise-free assumption, the final generated image strictly satisfies the data fidelity condition: ∥y−A ˆx∥2 2 =0. Since A 𝑃𝑁 ≡ 0, all prior modulation induced by the epistemic uncertainty 𝚺𝑡 is strictly constrained within the null space."

    The zero-residual result is exactly the definition of Eq. (20), where the estimate is constructed as x̂ = A†y + (I−A†A)x̄. Left-multiplying by A gives AA†y + A(I−A†A)x̄ = y + 0 by the Moore-Penrose identities, with no use of Σ_t, F_t, or UCPM. Thus 'strict data consistency' is an algebraic consequence of the update's form, not a consequence of the proposed uncertainty mechanism. The proposition restates its own construction and cannot fail, so the claimed theoretical guarantee is self-definitional rather than an independent result.

  2. other [Section 4.4, Eq. (17)-(19); Section 4.5, Proposition 4.2; Appendix B.2]
    "the error induced by a step of size Δ𝑡 is bounded as: 𝛿(Δ𝑡)≈(Δ𝑡)𝜌·𝑈𝑡≤𝐵 ... By solving this error constraint, the maximum admissible step size Δ𝑡Active can be derived ... Assuming the drift field is 𝐿-Lipschitz continuous and SATP bounds the single-step truncation error within the budget 𝐵. Based on Grönwall's inequality, the global accumulated error possesses a deterministic analytical upper bound: ∥x𝐷𝑒𝑛𝑠𝑒−x𝐿𝐸𝐴𝐷𝑒𝑟∥2≤𝐶·𝐵·(𝑒𝐿𝑇−1)/𝐿."

    Eqs. (18)-(19) define the SATP step size by imposing δ(Δt) ≈ (Δt)^ρ U_t ≤ B and solving for ΔtActive, with B as an input budget. Proposition 4.2 then assumes exactly that 'SATP bounds the single-step truncation error within the budget B' and derives the standard Gronwall accumulation bound. The error bound therefore follows from an assumption that is the design goal of the step-size formula, not from a verified property of the schedule. Moreover, the proportionality in Eq. (17) and the approximation δ ≈ (Δt)^ρ U_t are asserted without proof, so the theorem never establishes that the implemented pruning rule satisfies its own premise. The advertised deterministic bound is thus circular in structure.

full rationale

The empirical part of LEADer is genuinely comparative: it is tested against external baselines (DDNM, DDPG, ProjDiff, SITCOM, PIRP, EquS) and the plug-and-play improvements are reported on standard benchmarks, so those results are not circular. There is no load-bearing self-citation chain or imported uniqueness theorem; the authors' own PIRP and EquS-related citations are used as experimental baselines, not as premises. The circularity is confined to the two theoretical propositions advertised as guarantees. Proposition 4.1 proves a property that is true by construction of the null-space update in Eq. (20), so it adds no independent support. Proposition 4.2 assumes the very local truncation bound that SATP's step-size rule is designed to enforce, so the Gronwall conclusion is conditional on an assumption equivalent to the construction. Separately, several non-circular gaps weaken the theoretical claims: Eq. (10) is introduced as a derived relationship but no derivation is given; Eq. (17) asserts proportionality between the second state derivative and Tr(Σ_t) without proof; and the prior Φ(x) in Eq. (11) is never instantiated. The implementation also appears infeasible as written, since F_t is an N×N Hessian and Eq. (15) requires an N×N inverse, while Table 4 reports 0.07-0.28% memory overhead, implying an undisclosed approximation. These issues are correctness and reproducibility concerns rather than circularity, so they do not raise the score beyond the partial circularity already present in the two propositions.

Assumptions & free parameters 4 free parameters · 7 assumptions · 0 invented entities

The central method relies on two unproved identities, Eq. (10) and Eq. (17), and on an unspecified prior Phi(x). The information-loss budget B is fitted to data, and the error-bound proof assumes the very local-error bound the algorithm is designed to enforce. These items carry most of the method's load.

free parameters (4)
  • B = 0.01 (default)
    Information loss budget in Eq. (18); selected by sensitivity analysis in Table 3 to balance quality and runtime.
  • eta = not reported
    Scaling factor in the step-size formula Eq. (19); its value affects the actual step sizes and whether the stated error bound can hold.
  • epsilon = not reported
    Numerical stability constant in Eq. (19) to avoid division by zero when U_t is small.
  • rho = 1 (set for DDIM)
    Local error order assumed in Eq. (18); the paper sets rho = 1 for the DDIM solver.
assumptions (7)
  • domain assumption Linear degradation model y = A x0 + n with known A and Gaussian noise n
    Standard zero-shot inverse problem setup; introduced in Eq. (1).
  • standard math Tweedie's formula gives the clean estimate x_0|t from the score function
    Used in Eq. (2) and Eq. (9); standard result for Gaussian diffusion.
  • ad hoc to paper Posterior covariance equals (1-alpha_bar_t)/alpha_bar_t times [I - (1-alpha_bar_t) F_t]
    Asserted via a second-order Taylor expansion in Section 4.2, Eq. (10), without proof; it is central to both UCPM and SATP.
  • ad hoc to paper The norm of the second state derivative is proportional to Tr(Sigma_t)
    Eq. (17) asserts this proportionality without derivation; it is the basis for the adaptive step-size formula.
  • domain assumption The drift field of the reverse ODE is L-Lipschitz
    Required for the global error bound in Proposition 4.2; not verified for learned diffusion denoisers.
  • ad hoc to paper A concrete prior Phi(x) with computable gradient and Hessian exists
    UCPM's update Eq. (14) depends on Phi, but the paper never specifies it.
  • domain assumption Noise-free observation y = A x0 for the strict consistency proof
    Proposition 4.1 assumes y lies in Range(A); experiments also include noisy cases with sigma_e = 0.05.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Local Epistemic Uncertainty Guided Active Sampling for Plug-and-play Diffusive Image Restoration." pith.science (2026). https://pith.science/paper/P6L6CI2F

@misc{pith2026260806981,
  author       = {Pith},
  title        = {Pith review of: Local Epistemic Uncertainty Guided Active Sampling for Plug-and-play Diffusive Image Restoration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P6L6CI2F}},
  note         = {Machine review of arXiv:2608.06981}
}
read the original abstract

Diffusion models have demonstrated remarkable effectiveness in image restoration tasks. However, when guiding image reconstruction, existing Diffusion Model-based Image Restoration (DMIR) methods typically rely on fixed data constraints and uniform step sizes, thereby overlooking the dynamic nature of the generative process. Such rigid designs render the models vulnerable to spatially non-uniform degradations, thus resulting in structural distortions and loss of fine details. Meanwhile, uniform step sizes introduce computational redundancy, whereas na\"ive step reduction strategies tend to accumulate approximation errors. To address these limitations, we propose a Local Epistemic Uncertainty Guided Active Sampling framework (LEADer). In the spatial domain, LEADer leverages pixel-wise uncertainty to dynamically modulate the prior strength within the null space, which effectively balances detail preservation and artifact suppression. In the temporal domain, it quantifies sampling stability via the uncertainty trace to enable adaptive trajectory pruning, thereby accelerating convergence. Theoretical proofs demonstrate that our framework achieves strict data consistency, while the trajectory pruning strategy admits a deterministic error bound, thereby guaranteeing stable convergence under skip sampling. Notably, our plug-and-play method can be seamlessly integrated into various DMIR baselines. Extensive experiments show that LEADer improves the performance of multiple state-of-the-art DMIR methods, while significantly reducing sampling time with negligible memory overhead. Code is available at https://github.com/JiaqiZhang-Sengoku/LEADer.

Figures

Figures reproduced from arXiv: 2608.06981 by the authors.

Figure 1
Figure 1. Qualitative comparisons of the proposed method [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Architecture of LEADer. Existing DMIR methods rely on fixed step-by-step sampling with uniform constraints, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparisons on five image restoration tasks. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of xˆ 0|𝑡 between DDNM and LEADer at different NFEs. significantly reducing LPIPS by 7.13% to 12.78%. Notably, unlike EquS [49], LEADer leverages a state-aware trajectory pruning strat￾egy to improve generation quality while achieving substantial sam￾pling s…
Figure 5
Figure 5. Figure 5: Ablation study on different sampling steps. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Comparison of PSNR, LPIPS, and sampling time [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Time-varying 𝑈𝑡 and Δ𝑡Active during SATP [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 58 canonical work pages

  1. [1]

    Yasar Utku Alçalar and Mehmet Akçakaya. 2024. Zero-Shot Adaptation for Approximate Posterior Sampling of Diffusion Models in Inverse Problems. In European Conference on Computer Vision (Lecture Notes in Computer Science, Vol. 15141). Springer, 444–460

    work page 2024

  2. [2]

    Ismail Alkhouri, Shijun Liang, Cheng-Han Huang, Jimmy Dai, Qing Qu, Saiprasad Ravishankar, and Rongrong Wang. 2025. SITCOM: Step-wise Triple-Consistent Diffusion Sampling For Inverse Problems. InInternational Conference on Machine Learning, Vol. 267. PMLR / OpenReview.net

    work page 2025

  3. [3]

    Yochai Blau and Tomer Michaeli. 2018. The Perception-Distortion Tradeoff. In IEEE Conference on Computer Vision and Pattern Recognition. Computer Vision Foundation / IEEE Computer Society, 6228–6237

    work page 2018

  4. [4]

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. 2022. Simple Baselines for Image Restoration. InEuropean Conference on Computer Vision (Lecture Notes in Computer Science, Vol. 13667). Springer, 17–33

    work page 2022

  5. [5]

    Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. 2023. Diffusion Posterior Sampling for General Noisy In- verse Problems. InInternational Conference on Learning Representations. OpenRe- view.net

    work page 2023

  6. [6]

    Hyungjin Chung, Jeongsol Kim, and Jong Chul Ye. 2023. Direct Diffusion Bridge using Data Consistency for Inverse Problems. InAnnual Conference on Neural Information Processing Systems

    work page 2023

  7. [7]

    Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. 2022. Improving Diffusion Models for Inverse Problems using Manifold Constraints. InAnnual Conference on Neural Information Processing Systems

    work page 2022

  8. [8]

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

Show all 70 references
  1. [9]

    Egiazarian

    Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen O. Egiazarian

  2. [10]

    Prafulla Dhariwal and Alexander Quinn Nichol. 2021. Diffusion Models Beat GANs on Image Synthesis. InAnnual Conference on Neural Information Processing Systems. 8780–8794

    2021

  3. [11]

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. 2016. Image Super-Resolution Using Deep Convolutional Networks.IEEE Transactions on Pattern Analysis and Machine Intelligence38, 2 (2016), 295–307

    2016

  4. [12]

    Bradley Efron. 2011. Tweedie’s formula and selection bias.J. Amer. Statist. Assoc. 106, 496 (2011), 1602–1614

    2011

  5. [13]

    Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. 2023. Generative Diffusion Prior for Unified Image Restoration and Enhancement. InIEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 9935–9946

    2023

  6. [14]

    Tomer Garber and Tom Tirer. 2024. Image Restoration by Denoising Diffusion Models with Iteratively Preconditioned Guidance. InIEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 25245–25254

    2024

  7. [15]

    Bingbing Gu, Saihua Cai, Jing Wang, Zhuole Li, Xiheng Jia, and Jiaqi Zhang

  8. [16]

    Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. 2014. Weighted Nuclear Norm Minimization with Application to Image Denoising. InIEEE Con- ference on Computer Vision and Pattern Recognition. IEEE Computer Society, 2862–2869

    2014

  9. [17]

    Linchao He, Hongyu Yan, Mengting Luo, Hongjie Wu, Kunming Luo, Wang Wang, Wenchao Du, Hu Chen, Hongyu Yang, Yi Zhang, et al . 2023. Fast and stable diffusion inverse solver with history gradient update.CoRRabs/2307.12070 (2023)

    2023 arXiv

  10. [18]

    Yuanjian He, Chen Zhang, Fasheng Chen, and Jiangbo Cao. 2026. Cinematte: Back- ground matting for virtual production and beyond.arXiv preprint arXiv:2605.18328 (2026)

    2026 arXiv

  11. [19]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. InAnnual Conference on Neural Information Processing Systems

    2020

  12. [20]

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. 2022. Elucidating the Design Space of Diffusion-Based Generative Models. InAnnual Conference on Neural Information Processing Systems

    2022

  13. [21]

    Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. 2022. Denois- ing Diffusion Restoration Models. InAnnual Conference on Neural Information Processing Systems

    2022

  14. [22]

    Guangchen Lan, Sipeng Zhang, Tianle Wang, Yuwei Zhang, Daoan Zhang, Xin- peng Wei, Xiaoman Pan, Hongming Zhang, Dong-Jun Han, and Christopher G Brinton. 2025. MaPPO: Maximum a Posteriori Preference Optimization with Prior Knowledge.arXiv preprint arXiv:2507.21183(2025)

    2025 arXiv

  15. [23]

    Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi

    Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunning- ham, Alejandro Acosta, Andrew P. Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. 2017. Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network. InIEEE C...

    2017

  16. [24]

    Alkhouri, Saiprasad Ravishankar, and Qing Qu

    Xiang Li, Soo Min Kwon, Ismail R. Alkhouri, Saiprasad Ravishankar, and Qing Qu. 2024. Decoupled Data Consistency with Diffusion Purification for Image Restoration.CoRRabs/2403.06054 (2024)

    2024 arXiv

  17. [25]

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. 2021. SwinIR: Image Restoration Using Swin Transformer. InIEEE/CVF International Conference on Computer Vision Workshops. IEEE, 1833–1844

    2021

  18. [26]

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. 2015. Deep Learning Face Attributes in the Wild. InInternational Conference on Computer Vision. IEEE Computer Society, 3730–3738

    2015

  19. [27]

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. 2022. DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps. InAnnual Conference on Neural Information Processing Systems

    2022

  20. [28]

    Andreas Lugmayr, Martin Danelljan, Andrés Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. 2022. RePaint: Inpainting using Denoising Diffusion Probabilis- tic Models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 11451–11461

    2022

  21. [29]

    Morteza Mardani, Jiaming Song, Jan Kautz, and Arash Vahdat. 2024. A Variational Perspective on Solving Inverse Problems with Diffusion Models. InInternational Conference on Learning Representations. OpenReview.net

    2024

  22. [30]

    McCann, Kyong Hwan Jin, and Michael Unser

    Michael T. McCann, Kyong Hwan Jin, and Michael Unser. 2017. Convolutional Neural Networks for Inverse Problems in Imaging: A Review.IEEE Signal Pro- cessing Magazine34, 6 (2017), 85–95

    2017

  23. [31]

    Mang Ning, Enver Sangineto, Angelo Porrello, Simone Calderara, and Rita Cuc- chiara. 2023. Input Perturbation Reduces Exposure Bias in Diffusion Models. In International Conference on Machine Learning (Proceedings of Machine Learning Research). PMLR, 26245–26265

    2023

  24. [32]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 10674– 10685

    2022

  25. [33]

    Bernstein, Alexander C

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. 2015. ImageNet Large Scale Visual Recognition Challenge.Int. J. Comput. Vis.115 (2015), 211–252

    2015

  26. [34]

    Weiss, Niru Maheswaranathan, and Surya Ganguli

    Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli

  27. [35]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2021. Denoising Diffusion Implicit Models. InInternational Conference on Learning Representations. Open- Review.net

    2021

  28. [36]

    Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. 2023. Pseudoinverse-Guided Diffusion Models for Inverse Problems. InInternational Conference on Learning Representations. OpenReview.net

    2023

  29. [37]

    Yang Song and Stefano Ermon. 2019. Generative Modeling by Estimating Gra- dients of the Data Distribution. InAnnual Conference on Neural Information Processing Systems. 11895–11907

    2019

  30. [38]

    Yang Song, Liyue Shen, Lei Xing, and Stefano Ermon. 2022. Solving Inverse Prob- lems in Medical Imaging with Score-Based Generative Models. InInternational Conference on Learning Representations. OpenReview.net

    2022

  31. [39]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2021. Score-Based Generative Modeling through Stochas- tic Differential Equations. InInternational Conference on Learning Representations. OpenReview.net

    2021

  32. [40]

    Charles M Stein. 1981. Estimation of the mean of a multivariate normal distribu- tion.The Annals of Statistics(1981), 1135–1151

    1981

  33. [41]

    Reihsmann, Alexander L

    Yicheng Tao, Fan Feng, Xin Luo, Conrad V. Reihsmann, Alexander L. Hopkirk, Jean-Philippe Cartailler, Marcela Brissova, Stephen C. J. Parker, Diane C. Saunders, and Jie Liu. 2024. Cntools: A Computational Toolbox for Cellular Neighborhood Analysis from Multiplexed Images. InPLo...

    2024

  34. [42]

    Darshan Thaker, Abhishek Goyal, and René Vidal. 2024. Frequency-Guided Posterior Sampling for Diffusion-Based Image Restoration.CoRRabs/2411.15295 (2024)

    2024

  35. [43]

    Kot, and Bihan Wen

    Chong Wang, Lanqing Guo, Zixuan Fu, Siyuan Yang, Hao Cheng, Alex C. Kot, and Bihan Wen. 2025. Reconciling Stochastic and Deterministic Strategies for Zero- shot Image Restoration using Diffusion Model in Dual. InIEEE/CVF Conference on Computer Vision and Pattern Recognition. C...

    2025

  36. [44]

    Ge Wang, Jong Chul Ye, Klaus Mueller, and Jeffrey A. Fessler. 2018. Image Reconstruction is a New Frontier of Machine Learning.IEEE Transactions on Medical Imaging37, 6 (2018), 1289–1296. MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Jiaqi Zhang, Zheng Pang, Rongrong Ga...

    2018

  37. [45]

    Yinhuai Wang, Jiwen Yu, and Jian Zhang. 2023. Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model. InInternational Conference on Learning Representations. OpenReview.net

    2023

  38. [46]

    Yi Wang, Lanling Zeng, Jiaqi Zhang, and Yang Yang. 2026. Zero-Shot Diffusive Image Restoration With Consistency.Signal Processing248 (2026), 110705

    2026

  39. [47]

    Yiqi Wang, Jiaqi Zhang, Taotao Cai, Zirui Liu, Qingqiang Sun, Zequn Sun, Zhangkai Wu, Manqing Dong, Mingkai Zheng, Xuefei Yin, and Yanming Zhu

  40. [48]

    Zhihao Wang, Jian Chen, and Steven C. H. Hoi. 2021. Deep Learning for Image Super-Resolution: A Survey.IEEE Transactions on Pattern Analysis and Machine Intelligence43, 10 (2021), 3365–3387

    2021

  41. [49]

    Kevin Zhou

    Chenxu Wu, Qingpeng Kong, Peiang Zhao, Wendi Yang, Wenxin Ma, Fenghe Tang, Zihang Jiang, and S. Kevin Zhou. 2025. Equivariant Sampling for Improving Diffusion Model-based Image Restoration.CoRRabs/2511.09965 (2025)

    2025

  42. [50]

    Hongjie Wu, Linchao He, Mingqin Zhang, Dongdong Chen, Kunming Luo, Mengt- ing Luo, Jizhe Zhou, Hu Chen, and Jiancheng Lv. 2024. Diffusion Posterior Proximal Sampling for Image Restoration. InACM International Conference on Multimedia. ACM, 214–223

    2024

  43. [51]

    Hongjie Wu, Mingqin Zhang, Linchao He, Ji-Zhe Zhou, and Jiancheng Lv. 2025. Enhancing Diffusion Model Stability for Image Restoration via Gradient Man- agement. InACM International Conference on Multimedia. ACM, 10768–10777

    2025

  44. [52]

    From Agent Traces to Trust: A Survey of Evidence Tracing and Execution Provenance in LLM Agents.CoRRabs/2606.04990 (2026)

    2026 arXiv

  45. [53]

    Yang Yang, Xi Zhang, Jiaqi Zhang, and Lanling Zeng. 2026. Parameterized image restoration with diffusion and gradient priors.Knowledge Based Systems338 (2026), 115488

    2026

  46. [54]

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. 2022. Restormer: Efficient Transformer for High- Resolution Image Restoration. InIEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 5718–5729

    2022

  47. [55]

    Bingliang Zhang, Wenda Chu, Julius Berner, Chenlin Meng, Anima Anandkumar, and Yang Song. 2025. Improving Diffusion Inverse Problem Solving with Decou- pled Noise Annealing. InIEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 20895–20905

    2025

  48. [56]

    Fan Zhang, Shijun Chen, and Hua Wang. 2026. L-Drive: Beyond a Single Mapping- Latent Context Drives Time Series Forecasting.arXiv preprint arXiv:2605.17730 (2026)

    2026 arXiv

  49. [57]

    Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. 2024. Diffusion Models: A Comprehensive Survey of Methods and Applications.Comput. Surveys56, 4 (2024), 105:1–105:39

    2024

  50. [58]

    Fan Zhang, Shiming Fan, and Hua Wang. 2026. What If We Let Forecasting Forget? A Sparse Bottleneck for Cross-Variable Dependencies. InForty-third International Conference on Machine Learning

    2026

  51. [59]

    Jiawei Zhang, Jiaxin Zhuang, Cheng Jin, Gen Li, and Yuantao Gu. 2024. Unleashing the Denoising Capability of Diffusion Prior for Solving Inverse Problems. In Annual Conference on Neural Information Processing Systems

    2024

  52. [60]

    Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Tim- ofte. 2022. Plug-and-Play Image Restoration With Deep Denoiser Prior.IEEE Transactions on Pattern Analysis and Machine Intelligence44, 10 (2022), 6360–6376

    2022

  53. [61]

    Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. 2017. Be- yond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising. IEEE Transactions on Image Processing26, 7 (2017), 3142–3155

    2017

  54. [62]

    Fan Zhang, Shiming Fan, and Hua Wang. 2026. Rethinking Activation Function: A Simple Path to Efficient and Accurate Time Series Forecasting.IEEE Transactions on Knowledge and Data Engineering(2026)

    2026

  55. [63]

    Yaozong Zheng, Bineng Zhong, Qihua Liang, Zhenjun Tang, Rongrong Ji, and Xianxian Li. 2022. Leveraging local and global cues for visual tracking via parallel interaction network.IEEE Transactions on Circuits and Systems for Video Technology33, 4 (2022), 1671–1683

    2022

  56. [64]

    Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. 2023. Denoising Diffusion Models for Plug-and-Play Image Restoration. InIEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 1219–1229. Local Epistemic Uncertainty ...

    2023

  57. [67]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang

  58. [2007]

    IEEE Transactions on Image Processing16, 8 (2007), 2080–2095

    Image Denoising by Sparse 3-D Transform-Domain Collaborative Filtering. IEEE Transactions on Image Processing16, 8 (2007), 2080–2095

    2007

  59. [2015]

    In International Conference on Machine Learning (JMLR Workshop and Conference Proceedings, Vol

    Deep Unsupervised Learning using Nonequilibrium Thermodynamics. In International Conference on Machine Learning (JMLR Workshop and Conference Proceedings, Vol. 37), Francis R. Bach and David M. Blei (Eds.). JMLR.org, 2256– 2265

  60. [2018]

    InIEEE/CVF Conference on Computer Vision and Pattern Recognition

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. InIEEE/CVF Conference on Computer Vision and Pattern Recognition. 586–595

  61. [2023]

    Diffusion Models in Vision: A Survey.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 9 (2023), 10850–10869

    2023

  62. [2026]

    DG-SCL: Diffusion-Guided Semantic Contrastive Learning for Imbalanced Malicious Traffic Detection.Information Sciences(2026), 123851

    2026

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.