Pith. sign in

REVIEW 3 major objections 3 minor 110 references

Learning Single Index Models with Diffusion Priors

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a diffusion model can recover a signal from measurements $y = f(Ax^*)$ without ever knowing or differentiating the link function $f$, by running one partial inversion and one unconditional sampling pass.

desk verdict The empirical recipe is strong, but Section 4's recovery guarantee rests on an unproven and actually false isotropic-Gaussian assumption about the measurement residual; the method deserves referee attention, the theory as written does not. read the letter →

arxiv 2505.21135 v1 pith:DZF2JXSB submitted 2025-05-27 cs.LG cs.CVstat.ML

Anqi Tang , Youming Chen , Shuchen Xue , Zhaoqiang Liu This is my paper
classification cs.LGcs.CVstat.ML
keywords diffusionmodelssingleindexinverseproblemspartialinversionunknownlinkfunction1-bitcompressedsensingscore-basedgenerativesignalrecovery
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that an unknown, possibly discontinuous nonlinear link in a single index model $y = f(Ax^*)$ can be bypassed entirely when recovering $x^*$ from a pre-trained diffusion model. The proposed route is to treat the empirical correlation $(1/m)A^\top y$ as a noisy version of $\mu x^*$, choose the diffusion time whose noise level matches the noise floor introduced by the link, and then run one partial inversion followed by one unconditional sampling pass. If correct, accurate image recovery under unknown nonlinearities would cost roughly 150 neural function evaluations, far fewer than posterior sampling methods that require knowing the link function and thousands of evaluations. The paper reports this behavior on FFHQ, ImageNet, and CIFAR-10 for 1-bit and cubic measurements.

What carries the argument

The load-bearing object is the partial inversion operator $G^\dagger_{t_*}$ followed by the unconditional generator $G$, started not at the clean-data time $\epsilon$ but at an intermediate time chosen by noise matching. The identity doing the work is Lemma 2's concentration bound, rewritten heuristically as $(1/(m\mu))A^\top y \approx x^* + (C'/(C\mu\sqrt{m}))\epsilon$, combined with the diffusion forward-process form $x_t = \alpha_t x_0 + \sigma_t \epsilon$. Equating the two noise scales fixes $t_*$ through $\sigma_{t_*}/\alpha_{t_*} = C_s/\sqrt{m}$. Theorem 3 supplies the numerical-analysis half: with a $k_2$-th order sampler and a $k_1$-th order inverter, the round trip $G \circ G^\dagger_t$ stays within $O(\sqrt{n}(h_{\max}^{k_2} + L h_{\max}^{k_1}))$ of the exact ODE flow, leaving only the question of whether the scaled measurement vector lies on that flow.

What would settle it

Run SIM-DMIS on a single index model at a fixed $m$ while sweeping $C_s$ around the predicted value $C_s = \sqrt{m}\,\sigma_{t_*}/\alpha_{t_*}$; the noise-matching mechanism predicts a clear optimum that shifts with $\sqrt{m}$. If the optimal $C_s$ does not track $\sqrt{m}$, or if reconstruction stays accurate even when the measurement noise has variance far exceeding $\sigma_{t_*}^2$, the Eq. (25) mechanism is not what drives the result. A second direct falsifier is phase retrieval with $f(x) = x^2$, where $\mu = 0$: the paper's route is meaningless there, so successful recovery on phase retrieval would indicate a different mechanism.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a diffusion model's sampling and inversion operators can act as a projector that maps the correlation estimate $(1/m)A^\top y$ onto the signal manifold, with no knowledge of $f$. Lemma 2 quantifies the estimate: with high probability, $\|(1/m)A^\top y - \mu x^*\|_\infty = O(\sqrt{\log(2n)}/\sqrt{m})$, so up to an unknown scale $\mu$ the measurement vector resembles the true signal contaminated by noise of order $1/\sqrt{m}$. The method starts the inversion at an intermediate time $t_*$ satisfying $\sigma_{t_*}/\alpha_{t_*} = C_s/\sqrt{m}$, matching that noise level, and returns $\hat{x} = G \circ G^\dagger_{t_*}(\alpha_{t_*} C'_s A^\top y/m)$. Theorem 3 then shows the round trip $G \circ G^\dagger_t$ nearly reproduces any point on a true ODE trajectory, with error $O(\sqrt{n}(h_{\max}^{k_2} + L h_{\max}^{k_1}))$; the paper argues, through the approximation in its Eq. (25), that the scaled correlation vector is such a point at $t_*$, so the output is close to $x^*$.

Load-bearing premise

The argument depends on the scaled measurement vector $\alpha_{t_*} C'_s A^\top y/m$ being nearly a draw from the diffusion model's noisy distribution at time $t_*$ along the true signal's trajectory, an approximate equality asserted in Eq. (25) rather than proved.

Editorial extensions

If this is right

  • For any single index link satisfying $\mu \neq 0$ and a finite fourth moment, the link function never needs to be evaluated or differentiated, so discontinuous links such as $\mathrm{sign}(\cdot)$ pose no extra obstacle.
  • Partial inversion is the key: full inversion from time $\epsilon$ treats $(1/m)A^\top y$ as a clean image and fails, while starting from $t_*$ treats it as a noisy image and succeeds, as the paper's SIM-DMFIS versus SIM-DMIS comparison shows.
  • The cost is one forward sampling pass plus one partial backward pass, around 150 neural function evaluations on 256-by-256 images, versus 1,000 or more for posterior sampling baselines and 11,555 for the quantized-compressed-sensing baseline.
  • Recovery still assumes the true signal lies in the generator's range, equivalently approximately in the diffusion model's data distribution, which is the standard generative-prior assumption for this line of work.

Reading between the lines

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

  • A fully rigorous recovery theorem would need a third term controlling the distance from the scaled correlation vector to the diffusion ODE trajectory and the distance from $x^*$ into the generator's range; the current Theorem 3 controls only the round-trip error after the vector is already on a trajectory.
  • The noise-matching schedule predicts a testable scaling law: the optimal tuning constant $C_s$ should track $\sqrt{m}$, and the optimal start time should shift later as measurements become scarcer; sweeping $m$ would test Eq. (26) directly.
  • Links with $\mu = 0$, such as phase retrieval with $f(x) = x^2$ or $f(x) = |x|$, lie outside the method's stated scope; using second- or higher-order moments might lift recovery from the scale ambiguity and extend the same partial-inversion idea.
  • The principle of beginning inversion at the noise level of a linear empirical estimator could transfer to other inverse problems where a statistic of the measurements is a noisy version of the latent, including non-Gaussian sensing matrices or blind settings.
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

3 major / 3 minor

Summary. The paper proposes SIM-DMIS, a non-iterative method for recovering a signal x* from single-index model measurements y = f(Ax*) with an unknown, possibly discontinuous link function f. The method forms the empirical linear projection (1/m)A^T y, rescales it by a tuned constant and an intermediate diffusion time t*, and applies a partial inversion followed by full sampling of a pre-trained diffusion ODE. The authors claim a theoretical guarantee: Lemma 2 shows sup-norm concentration of the projection around mu x*, Eq. (25) heuristically models the projection as x* plus isotropic Gaussian noise, and Theorem 3 bounds the discretization error of inversion-plus-sampling along exact ODE trajectories. Experiments on FFHQ, ImageNet, and CIFAR-10 with 1-bit and cubic measurements report improved reconstruction quality over DPS, DAPS, and QCS-SGM at substantially lower NFE.

Significance. If the theoretical claim were established, this would be a practically relevant result: it would give a fast, link-function-free recovery method for single index models with diffusion priors, and the experimental evidence (roughly 19.9 dB PSNR on FFHQ 1-bit with 150 NFEs versus 16.6 dB for DAPS-N with 1000 NFEs) is promising. The paper also contains a self-contained proof of a numerical discretization bound (Theorem 3) and standard concentration lemmas. However, as detailed below, the advertised recovery guarantee is not proven: the key connection between the measurements and the diffusion ODE trajectory is only asserted heuristically and is in fact violated by the anisotropic residual of the linear projection.

major comments (3)
  1. [Section 4, Eqs. (25)-(27) and text before Theorem 3] The recovery claim for SIM-DMIS relies on the assertion that alpha_{t*} C'_s A^T y/m can be approximately written as alpha_{t*} x* + sigma_{t*} epsilon with epsilon ~ N(0, I_n). This is never proved. Lemma 2 only provides sup-norm concentration ||(1/m)A^T y - mu x*||_infinity <= C' sqrt(log(2n))/sqrt(m), which does not imply that the residual is an isotropic independent Gaussian vector. A direct computation shows the residual has covariance with eigenvalue Var(f(g)g)/(m mu^2) along the x* direction and E[f(g)^2]/(m mu^2) in every orthogonal direction; for f = sign these are approximately 0.57/m and 1.57/m, respectively, and the entries share common randomness from the y_i. Consequently, the input to G-dagger_{t*} is not a sample of q_{t*}, and the hypothesis of Theorem 3 is not satisfied.
  2. [Section 4, Theorem 3] Theorem 3 (Eq. (29)) bounds the distance between G composed with G-dagger_t(bar{x}_t) and the analytic solution bar{x}_epsilon only for an input bar{x}_t that already lies on the ODE trajectory. It says nothing about the measurement vector (1/m)A^T y or about x*. The text before Theorem 3 bridges this gap with an 'if' condition that is neither derived nor empirically validated with theory; the claim that 'x-hat is close to x* under appropriate conditions' is therefore not a consequence of the theorem. Section 4 does not provide a recovery guarantee for SIM-DMIS, despite the abstract and introduction advertising one.
  3. [Section 3, Eqs. (25)-(26) and Section 1.2] The method's choice of t* is based on a heuristic rather than a derived quantity: Eq. (26) sets sigma_{t*}/alpha_{t*} = C_s/sqrt(m) with a tuned constant C_s, and the paper explicitly states in Section 1.2 that the partial-inversion idea is 'based on heuristic theoretical results.' Even if the residual were Gaussian, its variance would involve the unknown mu and the anisotropic covariance, so the tuned constants C_s and C'_s absorb the mismatch. This means the theoretical analysis has no predictive content for the actual noise structure, and the empirical success is the only evidence for the method. The paper should either provide a correct analysis of the anisotropic residual or clearly separate the heuristic from the proven statements.
minor comments (3)
  1. [Appendix B.1] The sentence 'from Lemma 4, from Lemma 4' contains a duplicated phrase.
  2. [Eq. (19)] The phrase 'G-dagger_t is the composition of of v_{j_t}, v_{j_t-1}, ..., v_1' contains a duplicated 'of'.
  3. [Section 5] The theory assumes ||x*||_2 = 1, but the experiments are run on images whose pixel values are not unit-norm; the paper should clarify how the identifiability normalization is handled or why the theory still applies.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the recovery guarantee is conditional on an unproven distributional heuristic, but no step reduces to its own inputs by construction.

full rationale

The claimed derivation chain is: Lemma 2 bounds \|(1/m)A^T y - \mu x^*\|_\infty; the paper then says this 'inspires us to express' (1/(m\mu))A^T y as x^* plus Gaussian noise (Eq. 25), chooses t^* via \sigma_{t^*}/\alpha_{t^*} = C_s/\sqrt{m} (Eq. 26), and feeds \alpha_{t^*} C'_s A^T y/m into G \circ G^\dagger_{t^*}. The text is explicit that this is a conditional statement: 'if the term \alpha_{t^*} C'_s A^T y/m in Eq. (27) can be approximately expressed as \alpha_{t^*} x^* + \sigma_{t^*} \epsilon ... then \hat{x} is close to x^*.' The antecedent is not derived from Lemma 2, since sup-norm closeness does not imply Gaussianity or isotropy, so the advertised recovery guarantee is incomplete; however, incompleteness is not circularity. The conclusion \hat{x} \approx x^* is not assumed in the derivation of Theorem 3, which is a self-contained discretization error bound for the inversion-sampling roundtrip along a fixed ODE trajectory. The constants C_s and C'_s are tunable hyperparameters chosen to match a heuristic, not parameters fitted to the target recovery and then relabeled as predictions. The citations to (Liu & Liu, 2022) supply standard SIM moment conditions and a concentration lemma used in the proof of Lemma 2; these do not embed the present paper's recovery claim, so the self-citation is not load-bearing in a circular sense. The main weakness, the unproven distributional matching step, is a correctness/completeness risk rather than a circularity.

Assumptions & free parameters 2 free parameters · 7 assumptions · 0 invented entities

The central claim depends on two tuned hyperparameters, several standard domain assumptions about the link function and generator range, and one ad hoc assumption that connects the measurement statistics to a diffusion trajectory. No new physical or model entities are introduced.

free parameters (2)
  • Cs = tuned on CIFAR ablation (range 50-60)
    Controls the intermediate start time t* via σ_t/α_t = Cs/√m (Eq. 26). Theoretically Cs should absorb unknown constants from Lemma 2, but no value is derived and reconstruction quality is sensitive to it.
  • C'_s = tuned on CIFAR ablation (favorable values around 50-55)
    Scales the input vector α_t* C'_s A^T y/m (Eq. 27). Accounts for unknown scale µ and other constants; lower values improve PSNR/SSIM in ablations.
assumptions (7)
  • domain assumption The link function f satisfies µ = E[f(a^T x*) a^T x*] ≠ 0
    Eq. (20); standard for SIM identifiability, excludes phase retrieval.
  • domain assumption Fourth moment of f(a^T x*) is finite
    Eq. (21); allows heavy-tailed links such as cubic.
  • domain assumption The ground truth x* lies in the range of the diffusion generator G
    Section 3, after Eq. (22); standard generative-prior assumption.
  • domain assumption The neural function xθ is Lipschitz continuous in its first argument
    Assumption 1, Section 4; needed for Lemma 3 and Theorem 3.
  • ad hoc to paper The scaled measurement vector α_t* C'_s A^T y/m is approximately a noised version of x* on the diffusion ODE trajectory
    Eq. (25) and the paragraph before Eq. (26); this heuristic connects Lemma 2 to Theorem 3 but is not proved.
  • standard math Regularity conditions on derivatives of the scaled data prediction function
    Appendix C; similar to Lu et al. 2022, needed for the ODE solver error bound.
  • domain assumption The pretrained diffusion model's score approximates the true score and x_ε is close to ground truth
    Section 4 text; used to say x̄_ε is close to x*. Not formalized.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Single Index Models with Diffusion Priors." pith.science (2026). https://pith.science/paper/DZF2JXSB

@misc{pith2026250521135,
  author       = {Pith},
  title        = {Pith review of: Learning Single Index Models with Diffusion Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DZF2JXSB}},
  note         = {Machine review of arXiv:2505.21135}
}
read the original abstract

Diffusion models (DMs) have demonstrated remarkable ability to generate diverse and high-quality images by efficiently modeling complex data distributions. They have also been explored as powerful generative priors for signal recovery, resulting in a substantial improvement in the quality of reconstructed signals. However, existing research on signal recovery with diffusion models either focuses on specific reconstruction problems or is unable to handle nonlinear measurement models with discontinuous or unknown link functions. In this work, we focus on using DMs to achieve accurate recovery from semi-parametric single index models, which encompass a variety of popular nonlinear models that may have {\em discontinuous} and {\em unknown} link functions. We propose an efficient reconstruction method that only requires one round of unconditional sampling and (partial) inversion of DMs. Theoretical analysis on the effectiveness of the proposed methods has been established under appropriate conditions. We perform numerical experiments on image datasets for different nonlinear measurement models. We observe that compared to competing methods, our approach can yield more accurate reconstructions while utilizing significantly fewer neural function evaluations.

Figures

Figures reproduced from arXiv: 2505.21135 by the authors.

Figure 1
Figure 1. An illustration of our three approaches. For SIM-DMS, we only perform the sampling from t ∗ to ϵ. For SIM-DMFIS, we perform the full inversion and sampling procedures. For SIM￾DMIS, we first perform the inversion from t ∗ to T, and then perform the sampling from T to ϵ. Algorithm 1 The SIM-DMIS approach Input: A ∈ R m×n, y ∈ R m, xθ, time steps ϵ = tN < tN−1 < . . . < t1 < t0 = T, generator G corresponding to the sa… view at source ↗
Figure 2
Figure 2. Examples of 1-bit reconstructed images for FFHQ [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Examples of cubic reconstructed images for FFHQ. Results for ImageNet (256×256) For ImageNet, we re￾port the results for m = n/16 = 12288. We utilize a pre￾trained conditional ImageNet 256×256 model along with its corresponding classifier sourced from ADM (Dhariwal & Nichol, 2021). We adhere to the default recommended configuration settings of these models, substituting the un￾conditional ImageNet 256×256 model init… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Examples of 1-bit reconstructed images for ImageNet [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Examples of cubic reconstructed images for ImageNet. 6. Conclusion In this paper, we present approaches for learning SIMs by making use of the sampling and inversion methods for pre-trained unconditional DMs. Theoretical analysis and numerical results are provided to i…
Figure 6
Figure 6. Figure 6: Examples of reconstructed images for CIFAR-10 with m = 1000 1-bit measurements. Results on CIFAR-10 (32×32) with Cubic Measurements The quantitative results are shown in [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Examples of reconstructed images for CIFAR-10 with m = 1000 cubic measurements. E. Additional Examples of Reconstructed Images for FFHQ and ImageNet with 1-bit Measurements In this section, we present some additional examples of reconstructed images for the FFHQ 256×25…
Figure 8
Figure 8. Figure 8: Examples of reconstructed images for FFHQ with m = n/8 = 24576 1-bit measurements. Top: Origin, Middle: SIM-DMS, Bottom: SIM-DMIS. 7 [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: Examples of reconstructed images for ImageNet with m = n/16 = 12288 1-bit measurements. 8 [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

110 extracted references · 67 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    G., and Tamir, J

    Aali, A., Daras, G., Levac, B., Kumar, S., Dimakis, A. G., and Tamir, J. I. Ambient diffusion posterior sampling: Solving inverse problems with diffusion models trained on corrupted data. https://arxiv.org/2403.08728, 2024

    arXiv 2024

  3. [3]

    M., Mardani, M., and Vasanawala, S

    Alkan, C., Oscanoa, J., Abraham, D., Gao, M., Nurdinova, A., Setsompop, K., Pauly, J. M., Mardani, M., and Vasanawala, S. Variational diffusion models for blind MRI inverse problems. In NeurIPS 2023 Workshop on Deep Learning and Inverse Problems, 2023

    2023

  4. [4]

    Invertible generative models for inverse problems: M itigating representation error and dataset bias

    Asim, M., Daniels, M., Leong, O., Ahmed, A., and Hand, P. Invertible generative models for inverse problems: M itigating representation error and dataset bias. In ICML, 2020

    2020

  5. [5]

    Exact asymptotics for phase retrieval and compressed sensing with random generative priors

    Aubin, B., Loureiro, B., Baker, A., et al. Exact asymptotics for phase retrieval and compressed sensing with random generative priors. In MSML, 2020

    2020

  6. [6]

    Bora, A., Jalal, A., Price, E., and Dimakis, A. G. Compressed sensing using generative models. In ICML, 2017

    2017

  7. [7]

    Bortoli, V. D. Convergence of denoising diffusion models under the manifold hypothesis. Transactions on Machine Learning Research, 2022. Expert Certification

    2022

  8. [8]

    Boufounos, P. T. and Baraniuk, R. G. 1-bit compressive sensing. In CISS, 2008

    2008

Show all 110 references
  1. [9]

    Candes, E. J. and Wakin, M. B. An introduction to compressive sampling. IEEE Signal Processing Magazine, 25 0 (2): 0 21--30, 2008. doi:10.1109/MSP.2007.914731

    2008

  2. [10]

    Monte C arlo guided denoising diffusion models for B ayesian linear inverse problems

    Cardoso, G., Le Corff, S., Moulines, E., et al. Monte C arlo guided denoising diffusion models for B ayesian linear inverse problems. In ICLR, 2023

    2023

  3. [11]

    Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions

    Chen, H., Lee, H., and Lu, J. Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions. In ICML, 2023 a

    2023

  4. [12]

    A unified framework for uniform signal recovery in nonlinear generative compressed sensing

    Chen, J., Scarlett, J., Ng, M., and Liu, Z. A unified framework for uniform signal recovery in nonlinear generative compressed sensing. In NeurIPS, 2023 b

    2023

  5. [13]

    Score approximation, estimation and distribution recovery of diffusion models on low-dimensional data

    Chen, M., Huang, K., Zhao, T., and Wang, M. Score approximation, estimation and distribution recovery of diffusion models on low-dimensional data. In ICML, 2023 c

    2023

  6. [14]

    Chen, S., Chewi, S., Li, J., Li, Y., Salim, A., and Zhang, A. R. Sampling is as easy as learning the score: T heory for diffusion models with minimal data assumptions. In ICLR, 2023 d

    2023

  7. [15]

    The probability flow ODE is provably fast

    Chen, S., Chewi, S., Lee, H., Li, Y., Lu, J., and Salim, A. The probability flow ODE is provably fast. In NeurIPS, 2024

    2024

  8. [16]

    Diffusion policy: Visuomotor policy learning via action diffusion

    Chi, C., Feng, S., Du, Y., Xu, Z., Cousineau, E., Burchfiel, B., and Song, S. Diffusion policy: Visuomotor policy learning via action diffusion. In Robotics: S cience and S ystems , 2023

    2023

  9. [17]

    Chung, H., Sim, B., Ryu, D., and Ye, J. C. Improving diffusion models for inverse problems using manifold constraints. In NeurIPS, 2022 a

    2022

  10. [18]

    Chung, H., Sim, B., and Ye, J. C. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In CVPR, 2022 b

    2022

  11. [19]

    Chung, H., Kim, J., Kim, S., and Ye, J. C. Parallel diffusion models of operator and image for blind inverse problems. In CVPR, 2023 a

    2023

  12. [20]

    T., Klasky, M

    Chung, H., Kim, J., Mccann, M. T., Klasky, M. L., and Ye, J. C. Diffusion posterior sampling for general noisy inverse problems. In ICLR, 2023 b

    2023

  13. [21]

    Chung, H., Lee, S., and Ye, J. C. Decomposed diffusion sampler for accelerating large-scale inverse problems. In ICLR, 2024

    2024

  14. [22]

    Intermediate layer optimization for inverse problems using deep generative models

    Daras, G., Dean, J., Jalal, A., and Dimakis, A. Intermediate layer optimization for inverse problems using deep generative models. In ICML, 2021

    2021

  15. [23]

    Score-guided intermediate level optimization: Fast L angevin mixing for inverse problems

    Daras, G., Dagan, Y., Dimakis, A., and Daskalakis, C. Score-guided intermediate level optimization: Fast L angevin mixing for inverse problems. In ICML, 2022

    2022

  16. [24]

    C., Milanfar, P., Dimakis, A

    Daras, G., Chung, H., Lai, C.-H., Mitsufuji, Y., Ye, J. C., Milanfar, P., Dimakis, A. G., and Delbracio, M. A survey on diffusion models for inverse problems. https://arxiv.org/2410.00083, 2024

    2024 arXiv

  17. [25]

    Image N et: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Image N et: A large-scale hierarchical image database. In CVPR, 2009

    2009

  18. [26]

    Modeling sparse deviations for compressed sensing using generative models

    Dhar, M., Grover, A., and Ermon, S. Modeling sparse deviations for compressed sensing using generative models. In ICML, 2018

    2018

  19. [27]

    and Nichol, A

    Dhariwal, P. and Nichol, A. Diffusion models beat GAN s on image synthesis. In NeurIPS, 2021

    2021

  20. [28]

    and Song, Y

    Dou, Z. and Song, Y. Diffusion posterior sampling for linear inverse problem solving: A filtering perspective. In ICLR, 2024

    2024

  21. [29]

    Dirac D iffusion: Denoising and incremental reconstruction with assured data-consistency

    Fabian, Z., Tinaz, B., and Soltanolkotabi, M. Dirac D iffusion: Denoising and incremental reconstruction with assured data-consistency. In ICML, 2023

    2023

  22. [30]

    Adapt and diffuse: Sample-adaptive reconstruction via latent diffusion models

    Fabian, Z., Tinaz, B., and Soltanolkotabi, M. Adapt and diffuse: Sample-adaptive reconstruction via latent diffusion models. Proceedings of machine learning research, 235: 0 12723, 2024

    2024

  23. [31]

    Generative diffusion prior for unified image restoration and enhancement

    Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., and Dai, B. Generative diffusion prior for unified image restoration and enhancement. In CVPR, 2023

    2023

  24. [32]

    Feng, B. T. and Bouman, K. L. Efficient B ayesian computational imaging with a surrogate score-based prior. In NeurIPS 2023 Workshop on Deep Learning and Inverse Problems, 2023

    2023

  25. [33]

    T., Smith, J., Rubinstein, M., Chang, H., Bouman, K

    Feng, B. T., Smith, J., Rubinstein, M., Chang, H., Bouman, K. L., and Freeman, W. T. Score-based diffusion models as principled priors for inverse imaging. In ICCV, 2023

    2023

  26. [34]

    and Rauhut, H

    Foucart, S. and Rauhut, H. A Mathematical Introduction to Compressive Sensing. Springer New York, 2013

    2013

  27. [35]

    Solving inverse problems with deep neural networks--robustness included? IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (1): 0 1119--1134, 2022

    Genzel, M., Macdonald, J., and M \"a rz, M. Solving inverse problems with deep neural networks--robustness included? IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (1): 0 1119--1134, 2022

    2022

  28. [36]

    Phase retrieval under a generative prior

    Hand, P., Leong, O., and Voroninski, V. Phase retrieval under a generative prior. In NeurIPS, 2018

    2018

  29. [37]

    G., and Levine, S

    Hansen-Estruch, P., Kostrikov, I., Janner, M., Kuba, J. G., and Levine, S. I DQL : I mplicit Q -learning as an actor-critic method with diffusion policies. https://arxiv.org/2304.10573, 2023

    2023 arXiv

  30. [38]

    and Hand, P

    Heckel, R. and Hand, P. Deep decoder: Concise image representations from untrained non-convolutional networks. In ICLR, 2019

    2019

  31. [39]

    Prompt-to- P rompt image editing with cross-attention control

    Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., and Cohen-Or, D. Prompt-to- P rompt image editing with cross-attention control. In ICLR, 2023

    2023

  32. [40]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. In NeurIPS, 2020

    2020

  33. [41]

    Y., Bae, H., and Chun, S

    Hong, S., Lee, K., Jeon, S. Y., Bae, H., and Chun, S. Y. On exact inversion of DPM-S olvers. In CVPR, 2024

    2024

  34. [42]

    Horowitz, J. L. Semiparametric and nonparametric methods in econometrics, volume 12. 2009

    2009

  35. [43]

    Hyder, R., Shah, V., Hegde, C., and Asif, M. S. Alternating phase projected gradient descent with generative priors for solving compressive phase retrieval. In ICASSP, 2019

    2019

  36. [44]

    and Hegde, C

    Jagatap, G. and Hegde, C. Algorithmic guarantees for inverse imaging with untrained network priors. In NeurIPS, volume 32, 2019

    2019

  37. [45]

    G., and Tamir, J

    Jalal, A., Arvinte, M., Daras, G., Price, E., Dimakis, A. G., and Tamir, J. Robust compressed sensing MRI with deep generative priors. In NeurIPS, 2021 a

    2021

  38. [46]

    Instance-optimal compressed sensing via posterior sampling

    Jalal, A., Karmalkar, S., Dimakis, A., and Price, E. Instance-optimal compressed sensing via posterior sampling. In ICML, 2021 b

    2021

  39. [47]

    A style-based generator architecture for generative adversarial networks

    Karras, T., Laine, S., and Aila, T. A style-based generator architecture for generative adversarial networks. In CVPR, 2019

    2019

  40. [48]

    Denoising diffusion restoration models

    Kawar, B., Elad, M., Ermon, S., and Song, J. Denoising diffusion restoration models. In NeurIPS, 2022

    2022

  41. [49]

    and Seelamantula, C

    Killedar, V. and Seelamantula, C. S. Compressive phase retrieval based on sparse latent generative priors. In ICASSP, 2022

    2022

  42. [50]

    Kim, G., Kwon, T., and Ye, J. C. Diffusionclip: Text-guided diffusion models for robust image manipulation. In CVPR, 2022

    2022

  43. [51]

    Variational diffusion models

    Kingma, D., Salimans, T., Poole, B., and Ho, J. Variational diffusion models. In NeurIPS, 2021

    2021

  44. [52]

    Diff W ave: A versatile diffusion model for audio synthesis

    Kong, Z., Ping, W., Huang, J., Zhao, K., and Catanzaro, B. Diff W ave: A versatile diffusion model for audio synthesis. In ICLR, 2021

    2021

  45. [53]

    and Hinton, G

    Krizhevsky, A. and Hinton, G. Learning multiple layers of features from tiny images. 2009

    2009

  46. [54]

    Laska, J. N. and Baraniuk, R. G. Regime change: B it-depth versus measurement-rate in compressive sensing. IEEE Trans. Sig. Proc., 60 0 (7): 0 3496--3505, 2012

    2012

  47. [55]

    Convergence for score-based generative modeling with polynomial complexity

    Lee, H., Lu, J., and Tan, Y. Convergence for score-based generative modeling with polynomial complexity. In NeurIPS, 2022

    2022

  48. [56]

    Towards faster non-asymptotic convergence for diffusion-based generative models

    Li, G., Wei, Y., Chen, Y., and Chi, Y. Towards faster non-asymptotic convergence for diffusion-based generative models. https://arxiv.org/2306.09251, 2023

    2023 arXiv

  49. [57]

    and Liu, Z

    Liu, J. and Liu, Z. Non-iterative recovery from nonlinear observations using generative models. In CVPR, 2022

    2022

  50. [58]

    and Scarlett, J

    Liu, Z. and Scarlett, J. The generalized L asso with nonlinear observations and generative priors. In NeurIPS, 2020 a

    2020

  51. [59]

    and Scarlett, J

    Liu, Z. and Scarlett, J. Information-theoretic lower bounds for compressive sensing with generative models. IEEE J. Sel. Areas Inf. Theory, 1 0 (1): 0 292--303, 2020 b

    2020

  52. [60]

    Sample complexity bounds for 1 -bit compressive sensing and binary stable embeddings with generative priors

    Liu, Z., Gomes, S., Tiwari, A., and Scarlett, J. Sample complexity bounds for 1 -bit compressive sensing and binary stable embeddings with generative priors. In ICML, 2020

    2020

  53. [61]

    Robust 1-bit compressive sensing with partial G aussian circulant matrices and generative priors

    Liu, Z., Ghosh, S., and Scarlett, J. Robust 1-bit compressive sensing with partial G aussian circulant matrices and generative priors. In ITW, 2021 a

    2021

  54. [62]

    Towards sample-optimal compressive phase retrieval with sparse and generative priors

    Liu, Z., Ghosh, S., and Scarlett, J. Towards sample-optimal compressive phase retrieval with sparse and generative priors. In NeurIPS, 2021 b

    2021

  55. [63]

    Generative principal component analysis

    Liu, Z., Liu, J., Ghosh, S., Han, J., and Scarlett, J. Generative principal component analysis. In ICLR, 2022 a

    2022

  56. [64]

    Misspecified phase retrieval with generative priors

    Liu, Z., Wang, X., and Liu, J. Misspecified phase retrieval with generative priors. In NeurIPS, 2022 b

    2022

  57. [65]

    Generalized eigenvalue problems with generative priors

    Liu, Z., Li, W., and Chen, J. Generalized eigenvalue problems with generative priors. In NeurIPS, 2024

    2024

  58. [66]

    D PM-S olver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps

    Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., and Zhu, J. D PM-S olver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022 a

    2022

  59. [67]

    D PM-S olver++: Fast solver for guided sampling of diffusion probabilistic models

    Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., and Zhu, J. D PM-S olver++: Fast solver for guided sampling of diffusion probabilistic models. https://arxiv.org/2211.01095, 2022 b

    2022 arXiv

  60. [68]

    and Kabashima, Y

    Meng, X. and Kabashima, Y. Quantized compressed sensing with score-based generative models. https://arxiv.org/2211.13006, 2022

    2022 arXiv

  61. [69]

    V., Jagatap, G., and Hegde, C

    Nguyen, T. V., Jagatap, G., and Hegde, C. Provable compressed sensing with generative priors via L angevin dynamics. https://arxiv.org/2102.12643, 2021

    2021 arXiv

  62. [70]

    A., et al

    Ongie, G., Jalal, A., Metzler, C. A., et al. Deep learning techniques for inverse problems in imaging. IEEE J. Sel. Areas Inf. Theory, 1 0 (1): 0 39--56, 2020

    2020

  63. [71]

    Effective real image editing with accelerated iterative diffusion inversion

    Pan, Z., Gherardi, R., Xie, X., and Huang, S. Effective real image editing with accelerated iterative diffusion inversion. In ICCV, 2023

    2023

  64. [72]

    Localizing object-level shape variations with text-to-image diffusion models

    Patashnik, O., Garibi, D., Azuri, I., Averbuch-Elor, H., and Cohen-Or, D. Localizing object-level shape variations with text-to-image diffusion models. In ICCV, 2023

    2023

  65. [73]

    and Vershynin, R

    Plan, Y. and Vershynin, R. The generalized L asso with non-linear observations. IEEE Trans. Inf. Theory, 62 0 (3): 0 1528--1537, 2016

    2016

  66. [74]

    Robust one-bit recovery via R e LU generative networks: Improved statistical rates and global landscape analysis

    Qiu, S., Wei, X., and Yang, Z. Robust one-bit recovery via R e LU generative networks: Improved statistical rates and global landscape analysis. In ICML, 2020

    2020

  67. [75]

    GAN -based projector for faster recovery with convergence guarantees in linear inverse problems

    Raj, A., Li, Y., and Bresler, Y. GAN -based projector for faster recovery with convergence guarantees in linear inverse problems. In ICCV, 2019

    2019

  68. [76]

    Multiscale structure guided diffusion for image deblurring

    Ren, M., Delbracio, M., Talebi, H., Gerig, G., and Milanfar, P. Multiscale structure guided diffusion for image deblurring. In ICCV, 2023

    2023

  69. [77]

    Solving linear inverse problems provably via posterior sampling with latent diffusion models

    Rout, L., Raoof, N., Daras, G., Caramanis, C., Dimakis, A., and Shakkottai, S. Solving linear inverse problems provably via posterior sampling with latent diffusion models. In NeurIPS, 2023

    2023

  70. [78]

    Palette: Image-to-image diffusion models

    Saharia, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., and Norouzi, M. Palette: Image-to-image diffusion models. In SIGGRAPH, 2022 a

    2022

  71. [79]

    L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E. L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022 b

    2022

  72. [80]

    J., and Norouzi, M

    Saharia, C., Ho, J., Chan, W., Salimans, T., Fleet, D. J., and Norouzi, M. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (4): 0 4713--4726, 2022 c

    2022

  73. [81]

    and Ho, J

    Salimans, T. and Ho, J. Progressive distillation for fast sampling of diffusion models. In ICLR, 2022

    2022

  74. [82]

    and Cevher, V

    Scarlett, J. and Cevher, V. An introductory guide to F ano's inequality with applications in statistical estimation. Information- T heoretic M ethods in D ata S cience , 2021

    2021

  75. [83]

    R., Hand, P., and Eldar, Y

    Scarlett, J., Heckel, R., Rodrigues, M. R., Hand, P., and Eldar, Y. C. Theoretical perspectives on deep learning methods in inverse problems. IEEE Journal on Selected Areas in Information Theory, 3 0 (3): 0 433--453, 2022

    2022

  76. [84]

    and Hegde, C

    Shah, V. and Hegde, C. Solving linear inverse problems using GAN priors: An algorithm with provable guarantees. In ICASSP, 2018

    2018

  77. [85]

    and Ahmed, A

    Shamshad, F. and Ahmed, A. Compressed sensing-based robust phase retrieval via deep generative priors. IEEE Sens. J., 21 0 (2): 0 2286--2298, 2020

    2020

  78. [86]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015

    2015

  79. [87]

    M., Zhang, Z., Hu, X., Qu, Q., and Shen, L

    Song, B., Kwon, S. M., Zhang, Z., Hu, X., Qu, Q., and Shen, L. Solving inverse problems with latent diffusion models via hard data consistency. In ICLR, 2024

    2024

  80. [88]

    Denoising diffusion implicit models

    Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. In ICLR, 2021 a

    2021

  81. [89]

    Pseudoinverse-guided diffusion models for inverse problems

    Song, J., Vahdat, A., Mardani, M., and Kautz, J. Pseudoinverse-guided diffusion models for inverse problems. In ICLR, 2023

    2023

  82. [90]

    and Ermon, S

    Song, Y. and Ermon, S. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019

    2019

  83. [91]

    P., Kumar, A., Ermon, S., and Poole, B

    Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. In ICLR, 2021 b

    2021

  84. [92]

    Dual diffusion implicit bridges for image-to-image translation

    Su, X., Song, J., Meng, C., and Ermon, S. Dual diffusion implicit bridges for image-to-image translation. ICLR, 2023

    2023

  85. [93]

    T., and Bouman, K

    Sun, Y., Wu, Z., Chen, Y., Feng, B. T., and Bouman, K. L. Provable probabilistic imaging using score-based generative priors. IEEE Transactions on Computational Imaging, 2024

    2024

  86. [94]

    Compressed sensing with deep image prior and learned regularization

    Van Veen, D., Jalal, A., Soltanolkotabi, M., et al. Compressed sensing with deep image prior and learned regularization. https://arxiv.org/1806.06438, 2018

    2018 arXiv

  87. [95]

    Wainwright, M. J. High-dimensional statistics: A non-asymptotic viewpoint, volume 48. Cambridge University Press, 2019

    2019

  88. [96]

    Edict: Exact diffusion inversion via coupled transformations

    Wallace, B., Gokul, A., and Naik, N. Edict: Exact diffusion inversion via coupled transformations. In CVPR, 2023

    2023

  89. [97]

    B ELM : Bidirectional explicit linear multi-step sampler for exact inversion in diffusion models

    Wang, F., Yin, H., Dong, Y.-J., Zhu, H., Zhang, C., Zhao, H., Qian, H., and Li, C. B ELM : Bidirectional explicit linear multi-step sampler for exact inversion in diffusion models. In NeurIPS, 2024

    2024

  90. [98]

    Zero-shot image restoration using denoising diffusion null-space model

    Wang, Y., Yu, J., and Zhang, J. Zero-shot image restoration using denoising diffusion null-space model. In ICLR, 2023

    2023

  91. [99]

    On the statistical rate of nonlinear recovery in generative models with heavy-tailed data

    Wei, X., Yang, Z., and Wang, Z. On the statistical rate of nonlinear recovery in generative models with heavy-tailed data. In ICML, 2019

    2019

  92. [100]

    Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust

    Wen, Y., Kirchenbauer, J., Geiping, J., and Goldstein, T. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. https://arxiv.org/2305.20030, 2023

    2023 arXiv

  93. [101]

    Whang, J., Lei, Q., and Dimakis, A. G. Compressed sensing with invertible generative models and dependent noise. In NeurIPS 2020 Workshop on Deep Learning and Inverse Problems, 2020

    2020

  94. [102]

    G., and Milanfar, P

    Whang, J., Delbracio, M., Talebi, H., Saharia, C., Dimakis, A. G., and Milanfar, P. Deblurring via stochastic refinement. In CVPR, 2022

    2022

  95. [103]

    Wu, L., Trippe, B., Naesseth, C., Blei, D., and Cunningham, J. P. Practical and asymptotically exact conditional sampling in diffusion models. In NeurIPS, 2023

    2023

  96. [104]

    Deep compressed sensing

    Wu, Y., Rosca, M., and Lillicrap, T. Deep compressed sensing. In ICML, 2019

    2019

  97. [105]

    Wu, Z., Sun, Y., Chen, Y., Zhang, B., Yue, Y., and Bouman, K. L. Principled probabilistic imaging using diffusion models as plug-and-play priors. In NeurIPS, 2024

    2024

  98. [106]

    Improving diffusion inverse problem solving with decoupled noise annealing

    Zhang, B., Chu, W., Berner, J., Meng, C., Anandkumar, A., and Song, Y. Improving diffusion inverse problem solving with decoupled noise annealing. arXiv preprint arXiv:2407.01521, 2024 a

    2024 arXiv

  99. [107]

    S., and Chang, S

    Zhang, G., Ji, J., Zhang, Y., Yu, M., Jaakkola, T. S., and Chang, S. Towards coherent image inpainting using denoising diffusion implicit models. In ICLR, 2023

    2023

  100. [108]

    P., and Kleijn, W

    Zhang, G., Lewis, J. P., and Kleijn, W. B. Exact diffusion inversion via bi-directional integration approximation. In ECCV, 2024 b

    2024

  101. [109]

    and Chen, Y

    Zhang, Q. and Chen, Y. Fast sampling of diffusion models with exponential integrator. In NeurIPS 2022 Workshop on Score-Based Methods, 2022

    2022

  102. [110]

    Uni PC : A unified predictor-corrector framework for fast sampling of diffusion models

    Zhao, W., Bai, L., Rao, Y., Zhou, J., and Lu, J. Uni PC : A unified predictor-corrector framework for fast sampling of diffusion models. In NeurIPS, 2024

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.