acceptodds
Under review as a conference paper at ICLR 2027

Diffusion Waltz: MCMC Posterior Sampling with Diffusion Priors in Three Steps

Abstract

Pretrained diffusion models are powerful priors for inverse problems, but posterior sampling under nonlinear, and possibly non-differentiable, forward models remains challenging. We introduce _diffusion waltz_, an MCMC method that uses SDEdit-style noising–denoising as a proposal, corrected via Metropolis–Hastings to sample exactly from the posterior. We show that this sampling mechanism strictly generalizes preconditioned Crank–Nicolson (pCN) to learned prior settings, and, like pCN, can sample from the posterior without ever evaluating the prior density. To improve mixing, we further introduce _conditional noising_, which injects observation information directly into the proposal while preserving exactness. Across a range of nonlinear inverse problems, with or without access to gradients, diffusion waltz matches or outperforms existing baselines in reconstruction accuracy and uncertainty calibration.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Diffusion models, Inverse problems, MCMC, Stochastic interpolants, Bayesian methods, High dimensional MCMC, Functional inverse problems, Generative priors

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_7MzaOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes Diffusion Waltz (DW), an MCMC framework for Bayesian inverse problems with learned diffusion or flow-matching priors. Its basic proposal combines forward noising with conditional denoising and uses a Metropolis–Hastings correction that does not require evaluating the prior density. With an exact, compatible conditional denoiser, the proposal is prior-reversible; in the Gaussian mirror setting, it recovers preconditioned Crank–Nicolson. Evaluation includes checkerboard and Gaussian-process examples, Navier–Stokes and Darcy inverse problems, and hierarchical 3D inverse design.

Critical strengths
  1. A coherent prior-density-free sampling framework. The prior-reversibility construction and the auxiliary-variable correction for conditional noising provide a useful mechanism for posterior sampling with implicit generative priors. The explicit pCN connection clarifies the method’s relationship to established Bayesian sampling algorithms.

  2. Meaningful controlled validation and relevant scientific applications. The exact-denoiser Gaussian-process experiment provides a valuable check of the ideal conditional-noising construction, including comparison with an analytically available posterior. The nonlinear PDE and 3D applications address practically relevant settings, including derivative-free inference and hierarchical generation.

Critical weaknesses
  1. Insufficient qualification of the denoising construction and endpoint treatment. Proposition 1 specifies a particular SDE realization of p(x0∣xt)p(x_0\mid x_t), not a unique conditional-sampling algorithm; the subsequent uniqueness claim therefore needs a precise scope. The treatment of the inner reverse process at zero noise, including any cutoff or final denoising step, is also unclear; clipping the outer noising level does not resolve this issue.

  2. Strong assumptions and unclear practical applicability of Theorem 2. The informal statement does not adequately convey the restrictive assumptions in Appendix B.2, particularly full-support positive densities, the strict condition Atαt<1A_t\alpha_t<1, far-field rejection control, and uniform control of averaged denoising errors. Their applicability to the learned priors and experiments is not established. The authors should align the two statements and provide verifiable sufficient conditions, while clarifying that fixed nonzero denoising error leaves a residual perturbation term rather than guaranteeing exact posterior recovery.

  3. Insufficient support for the mixing advantage of conditional noising. Section 3.2 motivates observation-informed proposals, but faster reduction of data misfit does not establish faster posterior mixing. There is no mixing-rate improvement or a corresponding perturbation bound for the evaluated conditional-noising algorithms. In particular, the acknowledged EKCN interaction bias and adaptive updates are not covered by the fixed-RWMH analysis. More convincing evidence of posterior exploration under matched computational budgets is needed.

  4. Insufficient differentiation from existing posterior samplers. Remark 1 and Section 4 acknowledge DAPS, PnP/split-Gibbs methods, and concurrent noising–denoising proposals, but the comparison remains too qualitative to assess incremental novelty. The paper should distinguish their target distributions, observation-conditioning mechanisms, correction steps, residual approximation biases, and computational requirements. Without this comparison, it is difficult to separate new contributions from mechanisms already used in related conditional-sampling frameworks.

  5. Limited empirical support for generality and practical advantage. Both PDE evaluations use only five test cases, while the 3D distributional metrics assess mass rather than the full geometry posterior. The absence of standard image inverse-problem evaluations using pretrained priors on datasets such as FFHQ or ImageNet also leaves broader applicability insufficiently tested. Additional validation or appropriately narrower claims would strengthen the empirical conclusions.

Questions for authors

See Critical Weaknesses

Reviewer_6P92OpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work focuses on posterior sampling using pretrained diffusion models as priors. The paper proposes an MCMC algorithm that combines noising and denoising as a proposal mechanism with Metropolis–Hastings correction. Further, the authors introduce an observation-informed proposal to improve mixing time. The method establishes a connection to pCN and provides two variants to address both differentiable and non-differentiable forward operators.

Critical strengths
  1. The paper tries to address issues with non-differentiable forward operators which is an area that deserves more attention in inverse problems.
  2. The writing is clear and sufficiently easy to understand.
  3. The MH proposals do not require evaluation of the prior densities.
Critical weaknesses
  1. In my opinion, the first two waltz steps i.e., SDEdit style renoising followed by denoising using the model are very well known steps (originating in the paper DAPS [1], FLOWER [2] among others). The MH correction step is essentially favoring moves to push the sampler towards regions of higher likelihoods, which DAPS does using Langevin MC. Overall, although the order of operations are a little different, the waltz algorithm itself is not entirely novel nor is it surprising.
  2. It is not clear how the initial sample x0∼p0x_0 \sim p_0 is obtained. Would it require a full run of the diffusion model to produce a clean sample? If this is true, I believe a lot of compute is used in running a model to generated an unconditional sample followed by actual posterior sampling steps. Kindly clarify how this initialization step is performed.
  3. Eq. 6 is reverting back to gradient based optimization followed by renoising (SDEdit) and denoising. This once again falls in the DAPS style algorithm category. Could the authors clarify whether there is a real conceptual distinction or is the bias argument in Remark 1 the only fundamental reason why Diffusion-waltz may be superior?
  4. The Kalman approach using multiple generation trajectories is similar to particle based SMC methods which are quite popular in the reward-tilting literature for conditional generation. The authors clarify that reversibility conditions do not hold in this variant. Therefore, it is not possible to ensure that the sampler actually samples the true posterior.
  5. The method is computationally slower than simpler guidance based methods (Tables 9 and 6) although performance improvements are visible (not across all metrics).
  6. Would it be possible to add experiments on images (e.g., FFHQ dataset)? This would clarify the comparisons in performance and compute against established baselines such as DPS, DAPS and others.

Overall: The diffusion waltz method appears useful, but in my opinion, is of incremental novelty and empirical evidence as shown by the experimental section do not adequately support the broad exactness and sampling-reliability claims shown mathematically.

References:

  1. Zhang, Bingliang, Wenda Chu, Julius Berner, Chenlin Meng, Anima Anandkumar, and Yang Song. "Improving diffusion inverse problem solving with decoupled noise annealing." In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 20895-20905. IEEE, 2025.
  2. Pourya, Mehrsa, Bassam El Rawas, and Michael Unser. "Flower: A flow-matching solver for inverse problems." In International Conference on Learning Representations, vol. 2026, pp. 91318-91343. 2026.
Questions for authors

Please see weaknesses section.

Reviewer_X6ATOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Diffusion Waltz, an MCMC sampler for inverse problems with diffusion priors. The proposal noises the current state and denoises it with the exact conditional sampler; a Metropolis-Hastings step then accepts or rejects it. Because the proposal is prior-reversible, the acceptance ratio needs only the likelihood. The paper shows equivalence to pCN under a Gaussian prior and proves a perturbation-stability result for the random-walk variant. It also introduces conditional noising, which uses a Levenberg-Marquardt or ensemble-Kalman step to inject observation information. Experiments cover a toy problem, GP regression, Navier-Stokes, Darcy flow, and 3D bracket design.

Critical strengths
  1. Simple and general construction. No prior density is needed, non-differentiable forward models are supported, and two-stage generators such as TRELLIS.2 are handled cleanly.
  2. Conditional noising substantially improves mixing over the random-walk variant and is competitive with strong baselines on hard Navier-Stokes settings.
  3. GP regression shows that the gradient-based variant can match exact posterior samples.
Critical weaknesses
  1. The best Navier-Stokes results come from an inexact variant. DW-EKCN builds its proposal from the current ensemble, which breaks reversibility, and the stability theory covers only the random-walk variant. The paper mentions an exact group-splitting alternative but provides no comparison with it.

  2. The PDE results support reconstruction quality more than posterior accuracy:

  • few test instances per benchmark;
  • the metrics are pointwise or aggregate and do not test spatial correlations or mode coverage;
  • chains are short, and acceptance becomes very low after convergence in the low-noise regimes;
  • there are no convergence diagnostics or long-run reference chains.
  1. Computational cost is not compared at a matched budget. On Darcy, the gains over DPS are modest while the runtime is substantially higher, and the paper reports no accuracy-versus-cost curves or ESS per unit cost.

  2. The 3D evaluation only checks mass, which is the observed quantity itself, so geometric diversity and mode coverage are untested.

  • The noise level used is high enough that the proposal is close to an independent prior draw, so the method may behave much like - rejection sampling, and its sampling efficiency is similar to ESS-Flow's.
  • The baselines require ad hoc two-stage modifications, which further confounds the comparison.
  1. The scope is narrower than claimed.
  • Gradient-based conditional noising scales cubically with observation dimension, and the scalable ensemble variant is the inexact one.
  • The PDE experiments are at moderate resolution, and the higher-resolution setting only reports acceptance rates.
  • The 3D experiment has a scalar observation.
  • No experiment has both a high-dimensional state and high-dimensional observations.
  1. The theory's relevance to the experiments is unclear. The paper does not check whether the key assumptions (denoiser-mean contraction and tail acceptance control) hold for the learned priors. Stable acceptance rates across resolutions do not by themselves show resolution-robust mixing or accuracy.

Minor issue: "DM-LM" in Appendix E.2.2 should be "DW-LM".

Questions for authors

The construction is elegant and conditional noising is useful in practice. However, the paper's core claim is accurate posterior sampling, and the evidence does not yet support it at the level claimed. The strongest PDE results come from an inexact variant, and the PDE and 3D evaluations do not test posterior structure beyond pointwise metrics and the observed scalar. Cost-matched comparisons are missing, and the regime where the method is clearly preferable still remains narrow.

1 comment

Dear Reviewer X6AT,

We greatly appreciate the reviewer's time and constructive feedbacks that have significantly improved the quality of the paper. During the discussion period, we have continued to strengthen the paper in our updated manuscript with new results that we believe addresses most of the reviewer's concerns. This includes:

  • Concrete comparisons of EKCN waltz with the exact "split" variant (Appendix D.1.1 in the updated manuscript)
  • Expanded PDE test cases from five to ten -- conclusions are broadly the same
  • Added a matched-budget comparison with DPS in the Darcy problem (page 45 in the updated manuscript)
  • Posterior evaluations on 3D geometry: we have updated our Table 3 to include Fréchet Distances (FD) on (a) Inception-v3 features of 2D snapshots of the 3D geometry, taken from 10 views, and (b) 3D PointBERT features (please see Table below). These new results show that diffusion waltz yield distribution results comparable to ESS-flow while substantially improving likelihood alignment.
  • A new FFHQ 256 experiment to demonstrate our methodology in a high state dimension + high observation dimension regime (Appendix F.3 in our updated manuscript)
  • A new Gaussian mixture benchmark to further validate our posterior accuracy claims (Appendix F.2.3 in our updated manuscript)
MethodInception FDPointBERT FD ×10²Mass bias (posterior sd)Mass sd ratioIn exact 90% interval
Exact (reference)13.591.90+0.010.960.90
Waltz14.501.97−0.021.030.92
ESS-Flow13.462.00+0.293.220.79
EnKG22.242.63+1.956.420.54
CFG32.377.32−0.1410.750.16
Prior29.755.12+22.2317.570.04

Considering these additions, we kindly ask whether you might consider revisiting your score. If any concerns remain, we would be happy to address them in the remaining time.

Best regards,

The Authors

Comments on the paper

4 comments

Dear Reviewers,

Thank you again for your reviews. The authors have posted their rebuttal. If you haven't already, please take a look and let them know whether it addresses your concerns. If anything is still unclear, feel free to ask follow-up questions while the discussion period is open.

Even a short note (e.g., whether you are keeping or updating your score, and why) would help both the authors and the final decision.

Thanks for your time!

Best, AC

Dear Authors: Thank you for your detailed responses to my questions and concerns. I think the explanations you have provided are sufficiently convincing for me to raise my scores especially for W1 and W3. The additional experimental results are helpful too. Please ensure all of these additional details and explanations are included in your final manuscript. Best wishes.

We thank all the reviewers for their valuable time and constructive feedbacks that have helped improve our paper significantly. We have now uploaded our revised manuscript that addresses most of the concerns raised in the reviews. The main changes are:

  • A "contributions" statement in the introduction to explicitly highlight our main contributions
  • An expanded Relevant works section (Section 4) that explains how we compare against other related works
  • In particular, the comparison with DAPS, which was formerly Remark 1, is now expanded into a full discussion in Appendix C.
  • New experiments and updated results:
    • Tables 1 and 2 on our PDE experiments now consider ten test examples instead of five
    • Table 3 on the 3D experiment now show 3D posterior accuracy, computed via Fréchet Distance (FD) using features obtained via InceptionV3 embeddings of 2D snapshots, as well as 3D features from PointBERT
    • New experiments on a Gaussian mixture benchmark to further demonstrate our posterior accuracy (Appendix F.2.3)
    • New experiments on FFHQ 256 for comparisons on a well-established image benchmark (Appendix F.3)
    • Comparison of EKCN with split-EKCN to justify the use of inexact EKCN (Appendix D.1.1)
    • Matched budget comparison of AdCN-waltz vs DPS on the Darcy inverse problem (Page 45)
  • An expanded Limitations section (Section 6) to highlight our current limitations and scope for future work.

We thank you for the engagements and we'd be grateful to know if these address your concerns and if there are anything else that needs addressing in the remaining time.

We thank the reviewer for their valuable feedbacks and reconsidering their score. Please find all the changes in our updated manuscript (changes in blue); in particular, to address W1+3, we have expanded the relevant works section and added a new Appendix C that highlights how we position against DAPS and other methods in the conditional denoising class. Please let us know if the updates are insufficient and we would be happy to revise further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.