acceptodds
Withdrawn from ICLR 2027

Probability Flow ODE driven diffusion posterior sampling for large scale inverse problem with two-stage measurement alignment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion-based posterior sampling combines learned diffusion priors with measurement likelihoods to solve inverse problems. However, existing methods typically evaluate measurement guidance using a predicted clean sample or posterior mean rather than the current diffusion state, which can introduce biased likelihood gradients under noisy observations and nonlinear forward operators. They also often require hundreds or thousands of reverse-time steps, resulting in expensive likelihood evaluations. We propose probability flow ODE posterior sampling (PF-ODE), a deterministic framework that applies measurement likelihood guidance at the current diffusion state directly into the sampler, eliminating the reliance on posterior mean estimate. This high-order probability flow ODE solver through Strang splitting works in guided sampling while reduces discretization error and likelihood evaluations. For challenging nonlinear inverse problems, we further introduce a two-stage guidance strategy: phase-based guidance first mitigates cycle skipping and establishes a robust initialization, followed by proximal optimization using full measurements to enforce data consistency and recover fine-scale structures. Experiments on noisy image inpainting, sparse-view CT, and seismic full-waveform inversion demonstrate improved reconstruction quality with fewer likelihood evaluations than existing diffusion posterior sampling methods.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
probability flow ODE, large scale inverse problem, Diffusion posterior sampling, proximal optimization, two-stage measurement alignment

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_EqYUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a probability-flow ODE-based diffusion posterior sampling framework combining Strang splitting with two-stage measurement guidance for inverse problems. The authors aim to address likelihood approximation bias and computational inefficiency through phase-based alignment and proximal data-fidelity updates. Experiments on image inpainting, sparse-view CT, and seismic full-waveform inversion are presented to demonstrate improvements in reconstruction quality and computational efficiency.

Critical strengths

None

Critical weaknesses

1) There are fundamental methodological inconsistencies, the novelty is unclear, and justification for the claimed contributions is insufficient. Specifically, the paper's central methodological claims are not adequately supported by its mathematical formulation. The authors repeatedly emphasize that the proposed framework eliminates reliance on posterior-mean-based measurement guidance, yet Eqs. (13) and (18)-(20) explicitly use the Tweedie posterior-mean estimate for data-fidelity updates. I could not find anything in the paper that establishes how these updates resolve the approximation bias that motivates the work. Furthermore, the probability-flow ODE in Eq. (22) appears inconsistent with the SDE in Eq. (21) without an explicitly defined reverse-time convention, while the proximal weighting in Eq. (14) appears to contradict its stated noise-dependent behavior (it is reversed). Algorithm 1 introduces additional inconsistencies with the main formulation, particularly in the definition of the phase-based guidance term. Beyond these mathematical issues, the methodological novelty is unclear to me. The proximal data-fidelity formulation in Section 4.1, CG-based corrections of denoised estimates, and operator-splitting approaches for accelerated guided diffusion are already well-established in the literature. The proposed two-stage strategy also relies on previously developed measurement-alignment techniques. The paper does not sufficiently distinguish its contributions from these existing approaches or explain why their combination constitutes a meaningful methodological advance. Moreover, the relationship between phase retrieval, dynamic time warping, low-frequency guidance, and the actual measurement operators is not clearly formulated.

2) Inadequate experimental validation, insufficient reproducibility, and unsubstantiated claims of reconstruction quality and computational efficiency. The experimental evaluation is unfortunately insufficient to support the paper's broad claims. First, fundamental experimental details are missing (e.g., the number of evaluated images or hyperparameter specifications). The natural-image experiments consider only inpainting, while the CT experiments use synthetic projections of fastMRI images rather than actual CT data. The nonlinear FWI evaluation is supported only by a single qualitative example, without a comprehensive quantitative assessment across test cases or acquisition conditions. Second, the comparisons do not convincingly establish superiority over existing methods. From table captions for both Tab.1 and 2, I suspect that the authors have reported PSNR/SSIM results for a single image which is not consistent with how these results are reported in literature. Furthermore, the baseline discussion is severely inadequate (e.g., what is DPS-ΠGDM?). The qualitative results in Fig. 5 also shows noise residuals for both (e) and (g) which are the proposed methods. Third, the central efficiency claims are not substantiated by systematic runtime or computational-cost comparisons. The proposed iterative proximal updates require additional forward and adjoint evaluations, yet these costs are not accounted for when claiming improvements from fewer diffusion steps. Finally, I can also see some inconsistencies in the reported LPIPS values between Tables 1 and 3. Together with incorrect or misleading figure captions, this raise further concerns about the reliability and clarity of the experimental evaluation.

Questions for authors
  1. How many images or test cases were evaluated for each task, and what is the variability of the reported results? Could the authors also clarify the implementation of the competing methods, particularly the baseline labeled DPS-ΠGDM? Original ΠGDM work is a separate method than DPS.

  2. Diffusion-based inverse problem solving is a highly saturated and rapidly evolving research area, with numerous advances in efficient posterior sampling introduced in recent years. However, the paper primarily motivates and evaluates its contributions against relatively outdated methods such as DPS and ΠGDM from 2023. Given the substantial progress in fast diffusion-based inverse solvers, could the authors clarify how their method advances beyond the current state of the art and why more recent, competitive approaches are not included in the experimental comparisons?

Reviewer_zuj6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors propose to address inverse problems with diffusion-based prior using two types of data-fidelity guidance switched at some time point. They also apply a high-order: 2 NFEs (neural function evaluations) per iteration of a deterministic sampler.

Critical strengths

It makes sense to use phase alignment in early steps for some specific tasks.

Critical weaknesses

Section 4.1 is not novel. It is similar to DiffPIR [1] and many other works. It is also computed based on the estimate of x0, which is not aligned with the authors' narrative that it is "a coarse and unreliable estimate".

Section 4.2 (phase alignment) seems like an application specific step and an incremental contribution. You need to elaborate on how Eq. 18 is solved. Why phase alignment is required for tasks like inpainting and sparse-CT (it seems more relevant for phase retrieval)? If you use it, as presented in Eq. 18, only for the "Seismic full-waveform inversion" --- then you need to improve the presentation.

(Wizadwongsa & Suwajanakorn, 2023) already applied Strang splitting algorithm for guided diffusion sampling. What exactly is your contribution compared to it? The motivation for this step is also insufficient. Why Strang splitting is preferred over other higher-order samplers? What is the motivation compared to acceleration methods that do not require more than one NFE per iteration?

You should cite and discuss existing works that already improve efficiency by significantly reducing the number of steps, such as [2] CM4IR (accelerates guidance in consistency and diffusion models).

No theoretical insights.

There are no quantitative results averaged over a test set. Only a single example!

The experiments should include comparison with more recent methods. I would also like to see a compare with DiffPIR (whose data-fidelity step is quite similar) also on inpainting and CT.

I would like to see comparison on standard test sets (not single images) and benchmarks, and also for popular tasks like super-resolution and motion deblurring. Though, already in Table 1 the method does seem to be better than the DPS-PiGDM variant.

The authors should report the required number of NFEs and runtime of each method. They should also analyze the sensitivity of the proposed method to hyperparameters (number of iterations, guidance scale, proximal weight, switch time t*, ...).

[1] DiffPIR: "Denoising Diffusion Models for Plug-and-Play Image Restoration", CVPRW 2023

[2] CM4IR: "Zero-Shot Image Restoration Using Few-Step Guidance of Consistency Models (and Beyond)", CVPR 2025

Questions for authors

n/a

Reviewer_7iELOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper combines a Strang-splitting probability-flow ODE solver with two-stage measurement guidance for diffusion-based inverse problems. It first uses phase or low-frequency information, then applies a proximal update using full measurements. Experiments cover image inpainting, sparse-view reconstruction, and seismic full-waveform inversion. The authors report improved reconstruction quality with fewer sampling steps and include an ablation of the guidance components.

Critical strengths

The two-stage strategy has a useful motivation. Addressing coarse alignment before fitting full measurements is a reasonable approach, particularly for seismic inversion, where incorrect waveform alignment can lead optimization toward a wrong solution. This gives the method a clear practical motivation. Some reported results are promising.

Critical weaknesses
  1. The main claim does not match the algorithm, and the method description is inconsistent. The paper claims to remove dependence on the posterior-mean estimate and eliminate the resulting guidance bias. However, Eq. (13) still uses x^0\hat{x}_0 as the reference for proximal optimization, and Eqs. (18)–(20) also compute guidance through this estimate. No derivation or targeted experiment establishes that the bias is removed. Furthermore, Eq. (20) uses a warped-data residual, while Algorithm 1, line 7 uses the displacement norm. These are different objectives. Neither can correct traveltime errors as written: after warping, the residual no longer contains the traveltime error, and the gradient of the displacement norm is zero when u⋆u^\star is held fixed. In addition, Eqs. 13 to 17 match the data-proximal step and update rule of DiffPIR, which the paper cites only as an FWI baseline. This makes the methodological novelty unclear.
  2. The experiments do not establish the claimed efficiency or isolate the main contributions. Fewer outer sampling steps do not necessarily mean less computation: each proximal update can require multiple conjugate-gradient or Gauss–Newton iterations. Runtime and total forward/adjoint evaluations are not reported. It is also unclear whether the reported metrics are dataset averages or individual examples. Finally, the ablation does not isolate the proposed solver while keeping guidance fixed. These gaps make it difficult to judge both the practical gains and their source.
Questions for authors
  1. Are Tables 1–3 reporting dataset averages or individual examples? Please provide sample counts and explain why the apparently corresponding methods have different LPIPS values in Table 1 and 3.
  2. What are the total runtime and forward/adjoint evaluation counts, including inner optimization? Does the proposed solver improve the quality–cost trade-off when the guidance strategy is held fixed?

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

Related papers

Discussion (0)

Sign in to comment.