Pith. sign in

REVIEW 3 major objections 5 minor 253 references

A coordinate-wise SNR-guided modification of the DDIM sampler provably converges to the true Bayesian posterior for noisy linear inverse problems, with no likelihood gradients or particle filters.

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 12:42 UTC pith:O245SBWX

load-bearing objection A promising DDIM-based posterior sampler with a real, likely repairable, gap in the central proof; worth a serious referee but not publication as-is. the 3 major comments →

arxiv 2607.19333 v1 pith:O245SBWX submitted 2026-07-21 cs.LG cs.AIstat.ML

Provable diffusion-based posterior sampling for linear inverse problems via DDIM

Yuchen Jiao , Na Li , Changxiao Cai , Yuxin Chen , Gen Li This is my paper
classification cs.LG cs.AIstat.ML MSC 62F1568T07
keywords diffusion modelsDDIMlinear inverse problemsposterior samplingposterior consistencysingular value decompositionsignal-to-noise ratioimage restoration
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 paper claims that posterior sampling for noisy linear inverse problems reduces to a simple, coordinate-wise SNR-guided modification of the standard DDIM sampler. Along each singular direction of the measurement operator, the algorithm follows the learned diffusion prior while the measurement SNR is low, and switches to a calibrated measurement-based predictor once the measurement SNR dominates the diffusion SNR. Under an exact diffusion predictor and bounded signal support, the paper proves that the output distribution of this Posterior-DDIM sampler converges in total variation to the true posterior p_{X0|Y} as the discretization grid refines and the terminal diffusion time approaches zero. A second theorem extends the guarantee to arbitrary nonzero singular values with a correction term. If correct, this converts Bayesian posterior sampling into a lightweight DDIM-style procedure.

Core claim

Posterior-DDIM works in the singular-value basis of the measurement matrix A. For each direction s, the algorithm computes the threshold time τ_s at which the diffusion SNR α_t/σ_t equals the observation SNR Σ_ss/σ. In the measurement-dominated regime (t ≥ τ_s), the sampler replaces the DDIM update with a measurement-injected auxiliary process that has the same marginal law as the forward diffusion projected onto v_s. In the prior-dominated regime (t < τ_s), it keeps the standard DDIM update for observed directions, and uses a larger stochasticity parameter for unobserved directions. Theorems 1 and 2 assert that, with exact data predictors, bounded support, and suitable scaling of the stocha

What carries the argument

The central mechanism is the SNR-guided partition of singular directions S_meas^t = {s : α_t σ < Σ_ss σ_t} versus S_prior^t, equivalently encoded by the threshold time τ_s at which diffusion SNR matches observation SNR. The load-bearing identity is that the auxiliary forward process ξ_{t,s} = α_t Σ_{ss}^{-1} u_s^T Y + sqrt(σ_t^2 − α_t^2 σ^2 Σ_{ss}^{-2}) Z_s has the same marginal distribution as v_s^T X_t, which licenses replacing the DDIM update by the measurement-injected update in measurement-dominated coordinates. Prior-dominated and unobserved coordinates evolve by standard DDIM dynamics, with a correction term in the general-singular-value case to account for conditioning on the measure

Load-bearing premise

In the proof of Theorem 1 (Appendix A), the load-bearing premise is that the auxiliary measurement-injected coordinate at the switching time τ_s equals the raw observation Σ_{ss}^{-1}u_s^T Y; the construction yields α_{τ_s}Σ_{ss}^{-1}u_s^T Y, so identifying the auxiliary process with conditioning on Y holds only when α_{τ_s}=1.

What would settle it

Simulate Posterior-DDIM with an exactly known prior (e.g., a Gaussian mixture) and a measurement matrix whose singular value Σ_ss satisfies α_{τ_s}<1 at the matching time, then compare the empirical marginal of v_s^T X̂_{t_M} to the true posterior marginal along v_s. If the total-variation distance fails to vanish as the grid refines, the threshold identification is the culprit.

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

If this is right

  • Posterior sampling for noisy linear inverse problems reduces to coordinate-wise DDIM updates, avoiding likelihood-gradient computations and particle filtering.
  • Along low-SNR and unobserved directions, following the learned prior is provably the correct update; measurements are only imposed where observation SNR exceeds diffusion SNR.
  • The guarantees cover arbitrary, possibly rank-deficient measurement operators (Theorem 2), so tasks such as inpainting, super-resolution, deblurring, and compressed sensing fall under one proof.
  • The asymptotic TV convergence requires the terminal time to tend to 0 and the log-SNR grid to refine; with 100 DDIM steps the empirical metrics saturate, consistent with the theorem's regime.

Where Pith is reading between the lines

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

  • A testable diagnostic is whether the proof's identification at the switching time requires α_{τ_s}=1: the update (23) gives v_s^T X̂_{τ_s} = α_{τ_s} Σ_{ss}^{-1} u_s^T Y with zero residual noise, which equals the raw observation Σ_{ss}^{-1}u_s^T Y only if α_{τ_s}=1.
  • The per-direction switching rule suggests a sharp phase transition: coordinates with Σ_ss/σ above a time-dependent threshold are effectively pinned by the measurement, which could yield per-direction posterior contraction rates under exact predictors.
  • A natural extension is to replace the exact data predictor with a learned score in a controlled Gaussian-mixture setting and measure the TV gap empirically, isolating the error contributed by the switching boundary.
  • The double-limit statement supplies no convergence rate; a non-asymptotic bound would require quantitative control of the KL terms in Lemmas 1–6, which the argument leaves as an open constant.

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

3 major / 5 minor

Summary. The paper proposes Posterior-DDIM, a DDIM-type sampler for linear inverse problems with diffusion priors. The method partitions singular directions of the measurement operator according to whether the observation SNR exceeds the diffusion SNR, applies a calibrated measurement-based update in measurement-dominated directions, and a standard DDIM update in prior-dominated or unobserved directions. The main theoretical claims (Theorems 1 and 2) assert asymptotic posterior correctness in total variation under exact diffusion predictors, bounded support, and suitable choices of the stochasticity parameters. The paper also reports image-restoration experiments (inpainting, super-resolution, deblurring, compressed sensing) comparing favorably with several existing diffusion-based solvers.

Significance. If the theorems are correct, the contribution is significant: it provides a simple, coordinate-wise modification of DDIM with a proof of posterior consistency for noisy linear inverse problems, at lower computational overhead than many alternatives. The algorithm is easy to implement, and the experimental results are broadly competitive. The paper also ships a detailed proof structure with auxiliary processes and DDIM convergence machinery. However, the central proof currently contains a load-bearing identification error that invalidates the stated theorems as written. The result is plausible and likely repairable, but the manuscript cannot be accepted until the proof is corrected.

major comments (3)
  1. [Appendix A, final paragraph; Eqs. (23), (27), (33)] At the threshold time τ_s, Eq. (23) gives v_s^T \hat X_{τ_s} = α_{τ_s} Σ_{ss}^{-1} u_s^T Y, because the residual variance vanishes when α_{τ_s}/σ_{τ_s}=Σ_{ss}/σ. The text instead asserts v_s^T \hat X_{τ_s}=Σ_{ss}^{-1}u_s^T Y, and that (X_{t_M}, v_s^T X_{τ_s}) has the same joint law as (X_{t_M}, Σ_{ss}^{-1}u_s^T Y). This is false unless α_{τ_s}=1. Since the proof conditions on this quantity to identify the posterior, Theorems 1 and 2 are not established as written. The gap is probably repairable: v_s^T X_{τ_s}/α_{τ_s} equals Σ_{ss}^{-1}u_s^T Y in distribution, and in the scalar Gaussian model the conditioning event is equivalent up to an invertible transformation. This normalization and the resulting conditional equivalence must be stated and proved explicitly.
  2. [Section 2, Eq. (9); Eq. (11); Appendix B, Eqs. (28)-(29)] The discretization parameter is defined as δ_{i+1}=λ_{t_i}-λ_{t_{i+1}}. But the SNR α_t/σ_t is assumed decreasing in t, and the sampler runs from t_0 (large, α→0, σ→1) to t_M→0, so for t_i>t_{i+1} we have λ_{t_i}<λ_{t_{i+1}} and hence δ_{i+1}<0. Many subsequent bounds in Appendix B explicitly require δ>0 (e.g., 1-e^{-2δ_{i+1}}≍δ_{i+1} in Eq. (29), and the estimates in Lemma 6). This is either a sign error in Eq. (9) or a fundamental inconsistency in the time indexing. The definition must be corrected to δ_{i+1}=λ_{t_{i+1}}-λ_{t_i} (or the time order reversed) and the proof checked with the corrected convention.
  3. [Appendix C, proof of Theorem 2] The proof of Theorem 2 is substantially terser than that of Theorem 1. In particular, Eq. (106) is asserted 'by a similar argument with Lemma 1' without providing the general-A analogue of the error-term analysis in Lemma 6, and the conditioning argument for the auxiliary sequence in Eq. (105) is only sketched. Since Theorem 2 is a central advertised contribution, the general-case proof needs to be either written out in full or reduced explicitly to Theorem 1 with verifiable assumptions. As it stands, the reader cannot verify that the correction term in Eq. (25) preserves the claimed convergence.
minor comments (5)
  1. [Appendix B.1, Lemma 5] The definition of \hat bσ_{t_i}^2 contains a typographical inconsistency: Eq. (50) writes σ_{t_i}^2-α_{t_i}^2σ^2Σ_{ss}^{-1}, while Eq. (23) and the surrounding text use σ^2Σ_{ss}^{-2}. Please correct the notation.
  2. [Section 4.1] The empirical evaluation finds the best performance at η0=0, whereas the theory requires η0=1 in Theorem 1 and η0=δ^{-p}→∞ in Theorem 2. The experimental configuration is therefore outside the proven regime. This mismatch should be acknowledged and discussed.
  3. [Section 3.1, Eqs. (19)-(20)] The partition S_meas^t and S_prior^t uses strict inequalities, but the threshold time τ_s is defined by equality. At the exact boundary, Algorithm 1 and the proof need a consistent convention; the current presentation leaves this ambiguous.
  4. [Theorem 1, Eq. (26)] The double limit lim_{t_M→0} lim_{δ→0} and the definition of δ, which includes α_{t_0} and |1-σ_{t_0}|, should be accompanied by a concrete example of a valid schedule, so the order of limits and the meaning of δ→0 are unambiguous.
  5. [Tables 1 and 2] The table formatting around the rows beginning '0 32' and '2 16' is corrupted; the hyperparameter values are not clearly aligned with columns.

Circularity Check

0 steps flagged

No significant circularity: the Appendix A issue is a proof gap, not a reduction by construction.

full rationale

The paper's central claim is an asymptotic TV-convergence theorem for a concrete coordinate-wise DDIM algorithm. The proof in Appendices A and C proceeds by constructing auxiliary sequences, bounding KL/TV errors, and invoking analytical lemmas; no parameter is fitted to data and then renamed as a prediction, and no 'uniqueness theorem' from the authors' prior work is used to force the conclusion. The self-citations in Section 1.3 (Li et al. 2024, Jiao et al. 2025, Cai et al. 2026, etc.) are background references on unconditional DDPM/DDIM convergence and are not load-bearing in the proof of Theorems 1 and 2. The one suspicious passage is Appendix A's final paragraph, which states v_s^T \hat X_{τ_s} = Σ_{ss}^{-1} u_s^T Y and asserts a joint-law equivalence; by the paper's own Eq. (23) the value is α_{τ_s} Σ_{ss}^{-1} u_s^T Y, so the displayed equality is false as written. This is a genuine mathematical error in the proof, apparently repairable because the threshold condition gives σ_{τ_s}/α_{τ_s} = σ/Σ_{ss}, making the scaled event equivalent to the observation equation. But it is a correctness gap, not a circular derivation: the posterior conclusion is not assumed as an input, the algorithm's measurement branch is not just a relabeling of the target, and the proof does not reduce to a self-citation or to a fitted quantity. I therefore find no circular step that can be exhibited under the stated criteria.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The main fragility is the auxiliary-process identification, which carries the posterior correctness claim. The free parameters η0 and η1 are tuned empirically and the optimal choices are outside the theoretical regime. No new physical or mathematical entities are introduced.

free parameters (1)
  • η0, η1 = η0=0, η1=16 in experiments; theory uses η0=1 or η0=η1=δ^{-p}
    DDIM stochasticity parameters are tuned empirically in Section 4.1. The theoretical choices required by Theorems 1-2 are not used in the experiments, so the empirical performance is not covered by the proven guarantee.
axioms (5)
  • domain assumption Exact diffusion score / data predictor access
    The theorems assume access to the true conditional expectations µ_t and ϵ_t, not learned approximations. Section 2 and Theorem 1 state this implicitly through the use of exact score functions.
  • domain assumption Bounded support of X0
    Theorem 1 states 'Assume that the support of X0 is bounded'; used throughout Appendix B to bound moments and score derivatives.
  • domain assumption Monotone continuous SNR schedule with α0/σ0=∞ and α∞/σ∞=0
    Section 2 assumes a strictly decreasing continuous α_t/σ_t, which guarantees well-defined threshold times τ_s.
  • domain assumption Known Gaussian measurement noise with variance σ²I
    The model (2) assumes ε∼N(0,σ²I) with known σ; Algorithm 1 uses σ explicitly.
  • ad hoc to paper Posterior identification equality v_s^T \hat X_{τ_s}=Σ^{-1}u_s^TY and the joint-law identity with X_{t_M}
    In Appendix A, the proof identifies the auxiliary process at the threshold time with the raw observation. This equality is not implied by Eq. (23) unless α_{τ_s}=1, and it is load-bearing for the claimed posterior convergence.

pith-pipeline@v1.3.0-alltime-deepseek · 48901 in / 16428 out tokens · 171220 ms · 2026-08-01T12:42:51.868374+00:00 · methodology

0 comments
read the original abstract

Diffusion-based methods have achieved remarkable empirical success in solving inverse problems. However, many existing posterior samplers either lack rigorous theoretical guarantees or incur substantial computational overhead. We propose a simple and efficient algorithm, called \pddim, for solving linear inverse problems with diffusion priors via a DDIM-type sampler. Our method requires only lightweight, coordinate-wise modifications to the standard DDIM update, while explicitly incorporating the measurement model. The key idea is to perform posterior sampling separately along each singular direction of the measurement operator: for each direction, the sampler follows the learned diffusion prior when the observation signal-to-noise ratio (SNR) is below the corresponding diffusion SNR, and switches to a calibrated measurement-based predictor otherwise. We prove that the proposed sampler converges to the Bayesian posterior conditioned on the measurements. Empirical results show that the proposed sampler performs favorably against existing diffusion-based posterior samplers across a range of image restoration tasks, achieving the best performance on the majority of evaluation metrics considered. Overall, our results convert posterior sampling for noisy linear inverse problems to simple coordinate-wise DDIM updates, yielding an efficient, easy-to-implement algorithm with provable posterior consistency.

Figures

Figures reproduced from arXiv: 2607.19333 by Changxiao Cai, Gen Li, Na Li, Yuchen Jiao, Yuxin Chen.

Figure 1
Figure 1. Figure 1: FID versus PSNR, SSIM, and LPIPS for different inpainting algorithms on the CelebA dataset. [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Effect of the NFE on the ImageNet super-resolution task. [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗

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

253 extracted references · 39 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2303.05754 , year=

    Fast Diffusion Sampler for Inverse Problems by Geometric Decomposition , author=. arXiv preprint arXiv:2303.05754 , year=

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:2410.13738 , year=

    Instance-dependent Convergence Theory for Diffusion Models , author=. arXiv preprint arXiv:2410.13738 , year=

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:2503.10103 , year=

    Improving Diffusion-based Inverse Algorithms under Few-Step Constraint via Learnable Linear Extrapolation , author=. arXiv preprint arXiv:2503.10103 , year=

    arXiv

  4. [4]

    arXiv preprint arXiv:2510.27562 , year=

    Optimal Convergence Analysis of DDPM for General Distributions , author=. arXiv preprint arXiv:2510.27562 , year=

    arXiv

  5. [5]

    arXiv preprint arXiv:2606.23627 , year=

    Diffusion Models Adapt to Low-Dimensional Structure Under Flexible Coefficient Choices , author=. arXiv preprint arXiv:2606.23627 , year=

    Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:2206.09012 , year=

    Diffusion models as plug-and-play priors , author=. arXiv preprint arXiv:2206.09012 , year=

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2512.08022 , year=

    Provable diffusion posterior sampling for Bayesian inversion , author=. arXiv preprint arXiv:2512.08022 , year=

    arXiv

  8. [8]

    Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Come-Closer-Diffuse-Faster: Accelerating Conditional Diffusion Models for Inverse Problems through Stochastic Contraction , author=. Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  9. [9]

    Advances in Neural Information Processing Systems , volume=

    Solving linear inverse problems provably via posterior sampling with latent diffusion models , author=. Advances in Neural Information Processing Systems , volume=

  10. [10]

    International conference on learning representations , year=

    Pseudoinverse-guided diffusion models for inverse problems , author=. International conference on learning representations , year=

  11. [11]

    International Conference on Learning Representations , year=

    Diffusion Posterior Sampling for General Noisy Inverse Problems , author=. International Conference on Learning Representations , year=

  12. [12]

    Advances in Neural Information Processing Systems , volume=

    Denoising Diffusion Restoration Models , author=. Advances in Neural Information Processing Systems , volume=

  13. [13]

    arXiv preprint arXiv:2410.00083 , year=

    A survey on diffusion models for inverse problems , author=. arXiv preprint arXiv:2410.00083 , year=

    Pith/arXiv arXiv

  14. [14]

    Park, Trevor and Casella, George , journal=. The. 2008 , publisher=

    2008

  15. [15]

    IEEE Transactions on signal processing , volume=

    Bayesian compressive sensing , author=. IEEE Transactions on signal processing , volume=. 2008 , publisher=

    2008

  16. [16]

    Communications of the ACM , volume=

    Exact matrix completion via convex optimization , author=. Communications of the ACM , volume=. 2012 , publisher=

    2012

  17. [17]

    Journal of the American Statistical Association , volume=

    The spike-and-slab lasso , author=. Journal of the American Statistical Association , volume=. 2018 , publisher=

    2018

  18. [18]

    Journal of the American statistical Association , volume=

    Variable selection via nonconcave penalized likelihood and its oracle properties , author=. Journal of the American statistical Association , volume=. 2001 , publisher=

    2001

  19. [19]

    Proceedings of the National Academy of Sciences , volume=

    Message-passing algorithms for compressed sensing , author=. Proceedings of the National Academy of Sciences , volume=. 2009 , publisher=

    2009

  20. [20]

    IEEE Transactions on image processing , volume=

    Bayesian compressive sensing using Laplace priors , author=. IEEE Transactions on image processing , volume=. 2009 , publisher=

    2009

  21. [21]

    International Conference on Learning Representations , year=

    Solving inverse problems in medical imaging with score-based generative models , author=. International Conference on Learning Representations , year=

  22. [22]

    arXiv preprint arXiv:2510.21890 , year=

    The principles of diffusion models , author=. arXiv preprint arXiv:2510.21890 , year=

    Pith/arXiv arXiv

  23. [23]

    IEEE Transactions on Signal Processing , volume=

    Nonconvex optimization meets low-rank matrix factorization: An overview , author=. IEEE Transactions on Signal Processing , volume=. 2019 , publisher=

    2019

  24. [24]

    Proceedings of the International Congress of Mathematicians, Seoul, South Korea , volume=

    Mathematics of sparsity (and a few other things) , author=. Proceedings of the International Congress of Mathematicians, Seoul, South Korea , volume=. 2014 , organization=

    2014

  25. [25]

    arXiv preprint arXiv:2506.03979 , year=

    Solving inverse problems via diffusion-based priors: An approximation-free ensemble sampling approach , author=. arXiv preprint arXiv:2506.03979 , year=

    Pith/arXiv arXiv

  26. [26]

    IEEE Transactions on Computational Imaging , volume=

    Provable probabilistic imaging using score-based generative priors , author=. IEEE Transactions on Computational Imaging , volume=. 2024 , publisher=

    2024

  27. [27]

    arXiv preprint arXiv:2405.15643 , year=

    An unconditional representation of the conditional score in infinite-dimensional linear inverse problems , author=. arXiv preprint arXiv:2405.15643 , year=

    Pith/arXiv arXiv

  28. [28]

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

    The unreasonable effectiveness of deep features as a perceptual metric , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  29. [29]

    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

  30. [30]

    arXiv preprint arXiv:2406.06959 , year=

    Unleashing the denoising capability of diffusion prior for solving inverse problems , author=. arXiv preprint arXiv:2406.06959 , year=

    Pith/arXiv arXiv

  31. [31]

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

    Deep learning face attributes in the wild , author=. Proceedings of the IEEE international conference on computer vision , pages=

  32. [32]

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

    A style-based generator architecture for generative adversarial networks , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  33. [33]

    International journal of computer vision , volume=

    Imagenet large scale visual recognition challenge , author=. International journal of computer vision , volume=. 2015 , publisher=

    2015

  34. [34]

    Advances in Neural Information Processing Systems , volume=

    Snips: Solving noisy inverse problems stochastically , author=. Advances in Neural Information Processing Systems , volume=

  35. [35]

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

    Repaint: Inpainting using denoising diffusion probabilistic models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  36. [36]

    2019 , publisher=

    High-dimensional statistics: A non-asymptotic viewpoint , author=. 2019 , publisher=

    2019

  37. [37]

    arXiv preprint arXiv:2403.07902 , year=

    DecompDiff: diffusion models with decomposed priors for structure-based drug design , author=. arXiv preprint arXiv:2403.07902 , year=

    Pith/arXiv arXiv

  38. [38]

    arXiv preprint arXiv:1506.03365 , Year =

    Yu, Fisher and Zhang, Yinda and Song, Shuran and Seff, Ari and Xiao, Jianxiong , Title =. arXiv preprint arXiv:1506.03365 , Year =

    Pith/arXiv arXiv

  39. [39]

    arXiv preprint arXiv:2506.24042 , year=

    Faster Diffusion Models via Higher-Order Approximation , author=. arXiv preprint arXiv:2506.24042 , year=

    Pith/arXiv arXiv

  40. [40]

    arXiv preprint arXiv:2405.16876 , year=

    Transfer learning for diffusion models , author=. arXiv preprint arXiv:2405.16876 , year=

    Pith/arXiv arXiv

  41. [41]

    Journal of Chemical Information and Modeling , volume=

    Diffusion models in de novo drug design , author=. Journal of Chemical Information and Modeling , volume=. 2024 , publisher=

    2024

  42. [42]

    Journal of Materials Chemistry A , volume=

    Inverse design of porous materials: a diffusion model approach , author=. Journal of Materials Chemistry A , volume=. 2024 , publisher=

    2024

  43. [43]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Matfuse: controllable material generation with diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  44. [44]

    arXiv preprint arXiv:2503.09583 , year=

    Minimax optimality of the probability flow ode for diffusion models , author=. arXiv preprint arXiv:2503.09583 , year=

    Pith/arXiv arXiv

  45. [45]

    arXiv preprint arXiv:1707.08819 , year=

    A downsampled variant of imagenet as an alternative to the cifar datasets , author=. arXiv preprint arXiv:1707.08819 , year=

    Pith/arXiv arXiv

  46. [46]

    Nature , volume=

    De novo design of protein structure and function with RFdiffusion , author =. Nature , volume=. 2023 , publisher=

    2023

  47. [47]

    arXiv preprint arXiv:2206.04119 , year=

    Diffusion probabilistic modeling of protein backbones in 3d for the motif-scaffolding problem , author=. arXiv preprint arXiv:2206.04119 , year=

    Pith/arXiv arXiv

  48. [48]

    arXiv preprint arXiv:2402.04997 , year=

    Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design , author=. arXiv preprint arXiv:2402.04997 , year=

    Pith/arXiv arXiv

  49. [49]

    arXiv preprint arXiv:2009.09761 , year=

    Diffwave: A versatile diffusion model for audio synthesis , author=. arXiv preprint arXiv:2009.09761 , year=

    Pith/arXiv arXiv 2009

  50. [50]

    Probability Theory and Related Fields , volume=

    Measure estimation on manifolds: an optimal transport approach , author=. Probability Theory and Related Fields , volume=. 2022 , publisher=

    2022

  51. [51]

    Learning multiple layers of features from tiny images , year =

    Krizhevsky, Alex and Hinton, Geoffrey , address =. Learning multiple layers of features from tiny images , year =

  52. [52]

    Advances in neural information processing systems , volume=

    Gans trained by a two time-scale update rule converge to a local nash equilibrium , author=. Advances in neural information processing systems , volume=

  53. [53]

    arXiv preprint arXiv:2112.10741 , year=

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models , author=. arXiv preprint arXiv:2112.10741 , year=

    Pith/arXiv arXiv

  54. [54]

    arXiv preprint arXiv:2307.08123 , year=

    Solving inverse problems with latent diffusion models via hard data consistency , author=. arXiv preprint arXiv:2307.08123 , year=

    Pith/arXiv arXiv

  55. [55]

    Medical physics , volume=

    The lung image database consortium (LIDC) and image database resource initiative (IDRI): a completed reference database of lung nodules on CT scans , author =. Medical physics , volume=. 2011 , publisher=

    2011

  56. [56]

    Medical physics , volume=

    Low-dose CT image and projection dataset , author=. Medical physics , volume=. 2021 , publisher=

    2021

  57. [57]

    IEEE transactions on medical imaging , volume=

    The multimodal brain tumor image segmentation benchmark (BRATS) , author=. IEEE transactions on medical imaging , volume=. 2014 , publisher=

    2014

  58. [58]

    arXiv preprint arXiv:2403.17042 , year=

    Provably robust score-based diffusion posterior sampling for plug-and-play image reconstruction , author=. arXiv preprint arXiv:2403.17042 , year=

    Pith/arXiv arXiv

  59. [59]

    International Conference on Machine Learning , pages=

    Loss-guided diffusion models for plug-and-play controllable generation , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  60. [60]

    Advances in Neural Information Processing Systems , volume=

    Practical and asymptotically exact conditional sampling in diffusion models , author=. Advances in Neural Information Processing Systems , volume=

  61. [61]

    arXiv preprint arXiv:2308.07983 , year=

    Monte Carlo guided diffusion for Bayesian linear inverse problems , author=. arXiv preprint arXiv:2308.07983 , year=

    Pith/arXiv arXiv

  62. [62]

    arXiv preprint arXiv:2305.04391 , year=

    A variational perspective on solving inverse problems with diffusion models , author=. arXiv preprint arXiv:2305.04391 , year=

    Pith/arXiv arXiv

  63. [63]

    arXiv preprint arXiv:2212.00490 , year=

    Zero-shot image restoration using denoising diffusion null-space model , author=. arXiv preprint arXiv:2212.00490 , year=

    Pith/arXiv arXiv

  64. [64]

    The Twelfth International Conference on Learning Representations , year=

    Diffusion posterior sampling for linear inverse problem solving: A filtering perspective , author=. The Twelfth International Conference on Learning Representations , year=

  65. [65]

    2013 IEEE global conference on signal and information processing , pages=

    Plug-and-play priors for model based reconstruction , author=. 2013 IEEE global conference on signal and information processing , pages=. 2013 , organization=

    2013

  66. [66]

    Advances in Neural Information Processing Systems , volume=

    Diffusion models as plug-and-play priors , author=. Advances in Neural Information Processing Systems , volume=

  67. [67]

    IEEE Transactions on Computational Imaging , volume=

    Plug-and-play priors for bright field electron tomography and sparse interpolation , author=. IEEE Transactions on Computational Imaging , volume=. 2016 , publisher=

    2016

  68. [68]

    IEEE Transactions on Computational Imaging , volume=

    Plug-and-play ADMM for image restoration: Fixed-point convergence and applications , author=. IEEE Transactions on Computational Imaging , volume=. 2016 , publisher=

    2016

  69. [69]

    Advances in Neural Information Processing Systems , volume=

    Robust compressed sensing mri with deep generative priors , author=. Advances in Neural Information Processing Systems , volume=

  70. [70]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Score-based diffusion models as principled priors for inverse imaging , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  71. [71]

    IEEE Transactions on Computational Imaging , volume=

    Plug-and-Play ADMM for Image Restoration: Fixed-Point Convergence and Applications , author=. IEEE Transactions on Computational Imaging , volume=. 2017 , publisher=

    2017

  72. [72]

    International Conference on Learning Representations , year=

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

  73. [73]

    Advances in Neural Information Processing Systems , volume=

    Dmplug: A plug-in method for solving inverse problems with diffusion models , author=. Advances in Neural Information Processing Systems , volume=

  74. [74]

    Information and Inference: A Journal of the IMA , volume=

    Adaptivity and convergence of probability flow ODEs in diffusion generative models , author=. Information and Inference: A Journal of the IMA , volume=. 2026 , publisher=

    2026

  75. [75]

    Journal of Machine Learning Research , volume=

    O (d/T) convergence theory for diffusion probabilistic models under minimal assumptions , author=. Journal of Machine Learning Research , volume=

  76. [76]

    2005 , publisher=

    Inverse problem theory and methods for model parameter estimation , author=. 2005 , publisher=

    2005

  77. [77]

    Communications on Pure and Applied Mathematics: A Journal Issued by the Courant Institute of Mathematical Sciences , volume=

    An iterative thresholding algorithm for linear inverse problems with a sparsity constraint , author=. Communications on Pure and Applied Mathematics: A Journal Issued by the Courant Institute of Mathematical Sciences , volume=. 2004 , publisher=

    2004

  78. [78]

    IEEE Transactions on information theory , volume=

    Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information , author=. IEEE Transactions on information theory , volume=. 2006 , publisher=

    2006

  79. [79]

    IEEE Transactions on information theory , volume=

    Compressed sensing , author=. IEEE Transactions on information theory , volume=. 2006 , publisher=

    2006

  80. [80]

    arXiv preprint arXiv:2207.12598 , year=

    Classifier-free diffusion guidance , author=. arXiv preprint arXiv:2207.12598 , year=

    Pith/arXiv arXiv

Showing first 80 references.