acceptodds
Withdrawn from ICLR 2027

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

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

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 **Posterior-DDIM**, 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. Our code is available at https://anonymous.4open.science/r/DDIM_inv-E3C1.

Primary area:
learning theory
Keywords:
diffusion models, linear inverse problems, DDIM, posterior sampling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BhfYOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates diffusion-based posterior sampling for linear inverse problems. The authors employ singular value decomposition (SVD) of the linear measurement operator and compare the measurement SNR with the diffusion SNR along different singular directions. Based on this comparison, different DDIM update rules are applied to generate posterior samples. The authors also establish asymptotic convergence to the true posterior distribution under certain assumptions.

Critical strengths

1.Clear presentation and motivation.

The paper is well structured, with clearly stated research objectives. The proposed algorithm is relatively easy to understand for readers familiar with diffusion-based inverse problems.

2.Theoretical analysis.

The paper provides theoretical analysis and establishes asymptotic posterior convergence under specific assumptions.

Critical weaknesses

Limited algorithmic novelty compared with DDRM.

Although the authors propose Posterior-DDIM with asymptotic posterior convergence guarantees, its algorithmic design is closely related to DDRM. In particular, both methods employ essentially the same SVD-based noise partitioning strategy, and their updates along measurement-dominated directions have identical marginal distributions under certain parameter choices. While the proposed method introduces modifications to the sampling updates, their necessity and practical contributions relative to DDRM are not sufficiently demonstrated.

Questions for authors

Could the authors provide a direct comparison between Equation (8) of DDRM and Equations (22)–(24) of this paper, clarifying the main algorithmic differences and their significance? I would also appreciate targeted ablation studies demonstrating whether these modifications lead to meaningful empirical improvements.

Reviewer_ueawOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes relying on the fact that the one can see the observations of a linear Gaussian inverse problem as the observation of a given direction (defined by the eigenvalues of the linear operator AA) along the forward process of the diffusion generative model. One can then use directly the diffusion of these coordinates to guide over the same directions when sampling from the backward process until the appropriate time, while simply using the diffusion model for the other parts of the image.

The paper proposes a bound for the overall error in the idealized case of a perfect score based generative model and uses this perfect scenario bound to provide a bound for the real case where one uses an approximation of the score / denoising network, via the triangle inequality and the Pinsker bound.

The paper provides some numerical validation over three cases: 50% random inpainting, 4x super resolution (main) and Gaussian debluring and compressed sensing.

Critical strengths

None

Critical weaknesses

First of all, the paper does not clear state that the idea of decomposing the space into "observed parts" SmeasS^{meas} and unnobserved part SpriorS^{prior} has at least been used by two other posterior sampling algorithms, namely [1], [2] and [3]. Furthermore, the actual decomposition and definition of the onset times for each observation is really close to [3], see for example section 2.1 of [3].

Furthermore, the bound given is only non vacuous in idealized cases. Indeed, while the fact that it holds for the idealized case is nothing new and is already shown in [3] (see discussion around eq (2.10)), the bound for the practical case rely on the triangle inequality for the TV and the Pinsker inequality to bound the error between the "true backward" process and one that is implemented using an actual score. But as those errors involve the KL and MSEs between true score and learned score, they depend at least linearly on the dimension and therefore one should expect in realistic cases the KL to be bigger than 1. Thus, this would yield an overall TV bound for the practical algorithm that is bigger than 1, which of course is true for any distribution as the TV is bounded by 1. (or by 2, depending on the definition of the total variation).

Lastly, the numerical comparison rely only on papers published before 2024, which are not consider state of the art in the domains. Furthermore, the inverse problems are rather easy. It is well known that random inpainting is not a challenging problem for diffusion models, as there are no "missing patches". Furthermore 4x super resolution is also rather simple. If the authors could show that the proposed algorithm is better than state of the art algorithms such as [4] and [5] in the same / harder inverse problems as the one they consider this would strength their claim for relevance of the current paper.

[1] Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., & Van Gool, L. (2022, June). Repaint: Inpainting using denoising diffusion probabilistic models. In 2022 IEEE/CVF conference on computer vision and pattern recognition (CVPR) (pp. 11451-11461). IEEE.

[2] Kawar, B., Elad, M., Ermon, S., & Song, J. (2022). Denoising diffusion restoration models. Advances in neural information processing systems, 35, 23593-23606.

[3] Cardoso, G., Janati, Y., Le Corff, S., & Moulines, E. (2024, May). Monte Carlo guided Denoising Diffusion models for Bayesian linear inverse problems. In International Conference on Learning Representations (Vol. 2024, pp. 44001-44037).

[4] Janati, Y., Moufad, B., Abou El Qassime, M., Durmus, A. O., Moulines, E., & Olsson, J. (2025, February). A Mixture-Based Framework for Guiding Diffusion Models. In ICML.

[5] Wu, Z., Sun, Y., Chen, Y., Zhang, B., Yue, Y., & Bouman, K. L. (2024). Principled probabilistic imaging using diffusion models as plug-and-play priors. Advances in Neural Information Processing Systems, 37, 118389-118427.

Questions for authors

How do the authors compare to [4] and [5] ? How do the method function on missing patch inpainting problems, such as those shown in [4] Figure 1 for example ? Why are there no images illustrating the method in action ? Can the authors provide such images for missing patch inpainting problems ?

Reviewer_L6XpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Posterior-DDIM for linear inverse problems with pretrained diffusion priors. The sampler compares measurement and diffusion SNRs along the singular directions of the observation operator and applies either measurement-based updates or DDIM updates, with an additional correction for unequal singular values. The paper claims asymptotic posterior consistency under exact predictions and evaluates the method on inpainting, super-resolution, Gaussian deblurring, and compressed sensing using CelebA and ImageNet.

Critical strengths

The method addresses a useful problem: incorporating linear measurements into a pretrained diffusion sampler without task-specific retraining or differentiation through the denoising network. The coordinate-wise construction is relatively simple when the observation operator has an accessible singular value decomposition.

Critical weaknesses
  1. The main posterior convergence claim is not established. Theorem 1 permits bounded atomic priors. Consider d=n=1d=n=1, A=1A=1, X0≡0X_0 \equiv 0, and positive measurement noise. The exact posterior is δ0\delta_0. Under a standard VP schedule with η0=1\eta_0=1, the final update in Eq. (23) adds nondegenerate Gaussian noise at any sufficiently small positive terminal time. Hence, for every finite discretization, the output assigns zero probability to 0, so its total variation distance from the posterior is one. The iterated limit in Eq. (25) therefore cannot be zero.

  2. There is also a separate issue in the proof for general singular values. Eq. (103) conditions on the threshold states Xτs,sX_{\tau_s,s}, but the definition of pkstcp_k^{stc} omits these variables. Eq. (105) and the subsequent KL estimates then use this different target. Once the measurement-dominated set is empty, that target becomes the unconditional diffusion marginal. The proof needs to retain the observation-related conditioning and justify convergence to the intended posterior. Both issues affect the paper’s central theoretical contribution.

  3. The experiments do not establish the posterior correctness of the practical sampler. Theorem 1 requires η0=1\eta_0=1 for equal nonzero singular values and η0=η1\eta_0=\eta_1 otherwise, with η1=δ−p\eta_1=\delta^{-p}. The experiments instead use η0=0\eta_0=0 and η1=16\eta_1=16. In the inspected version of the released sampler, I could identify ordinary DDIM updates and measurement replacement, but not the threshold-state storage or the correction in Eq. (24). The implementation used for the Gaussian deblurring results therefore needs clarification. The reported PSNR, SSIM, LPIPS, and FID results assess restoration quality, but do not determine whether the conditional sampling distribution is correct. Averaging four outputs also does not verify posterior uncertainty. No tractable-posterior experiment is provided to establish the relationship between the theoretical construction, the practical implementation, and the claimed posterior distribution.

Questions for authors
  1. How does Theorem 1 handle the bounded atomic prior described above? If additional regularity assumptions are intended, please state them and justify the terminal-time TV limit under those assumptions.

  2. How are the threshold-state conditions in Eq. (103) retained in the target defined on page 36 and in the subsequent KL argument? Please identify the conditional density targeted by the corrected dynamics.

  3. Which implementation generated the reported tables, and does it include Eq. (24)? What convergence claim applies to the default setting η0=0,η1=16\eta_0=0,\eta_1=16? A comparison of the theoretical and practical variants on a known posterior, including mean and covariance errors as the discretization is refined, would materially affect my assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.