acceptodds
Withdrawn from ICLR 2027

FlowADMM: Plug-and-play ADMM with Flow-based Renoise-Denoise Priors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Plug-and-play (PnP) methods for solving inverse problems have recently achieved strong performance by leveraging denoising priors based on powerful generative diffusion and flow models. However, existing diffusion- and flow-based PnP methods typically rely on stochastic renoise-denoise operations, which complicate the analysis of their convergence behavior. In this work, we identify and formalize the deterministic renoise-denoise operator underlying flow-based plug-and-play methods. This perspective reveals that these methods implicitly define a deterministic operator given by the expectation of a denoiser over the latent noise distribution. Building on this insight, we propose FlowADMM, a PnP algorithm that integrates the renoise-denoise operator into the classical alternating direction method of multipliers (ADMM) framework. We establish convergence guarantees for FlowADMM under Lipschitz conditions on the underlying flow network, and extend the analysis to non-stationary time schedules. Empirically, FlowADMM matches or surpasses state-of-the-art performance among flow-based PnP methods on a range of inverse problems, including denoising, deblurring, super-resolution, and inpainting, while requiring fewer data consistency evaluations than prior approaches.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
inverse problems, optimization, computer vision, generative priors

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_WF7YOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper uses a PnP framework for image regularization, similar to PnP Flow Matching. In contrast to PnP Flow the authors use an ADMM framework instead of the forward backward splitting. Further they provide similar analysis for an averaged operator, which averages the denoising step for several Monte Carlo Samples z. The algorithm outperforms PnP Flow and Flower on standard image restoration tasks on CelebA and Cats.

Critical strengths
  1. The performance is genuinely strong. There are for instance strong gains in superresolution or random inpainting on CelebA.

  2. The paper is nicely written and justifies their algorithm somewhat theoretically.

Critical weaknesses
  1. First of all the averaged operator is alreadz done in PnP Flow at least empirically. They average over 5 samples, so defining the operator itself is not novel. Further, the convergence statements obtained are weak in practice.

  2. Proposition 1 is weak since it essentially requires the operator Sˉt−I\bar{S}_t - I to be 1 Lipschitz. This is, for a trained velocity field, on highly multimodal data essentially unavailable.

  3. The convergence statement is standard in Proposition 2, similar to PnP Flow and does not address to what point it converges to. This would be genuinely interesting, because with Liopschitz velocity and the (1-t) multiplication it is to no surprise that the algo converges.

  4. I am wondering where the gain in the numerical section really stems from. Since they optimize a bit different, I would love to see an ablation on PnP Flow and ADMM with the same grid. Because otherwise i do not concur with the thesis that ADMM is the factor that separates the methods, it might as well by hyperparameter optimization. I am aware that the late schedule has been applied to PnP Flow, but i would like to see a proper comparison with the same amount of hyperparams, same tuning budget to be fully convinced.

  5. There is already work unifying PnP methods and ADMM with flow based denoisers, see [1]. They in particular have in depth theoretical analysis although they do not share the renoise denoise framework, so a comparison with those algorithms (and the same pretrained velocities) would strengthen the submission.

Overall, I find the novelty to be very limited with respect to prior work and the outperformance of empirical results cannot be clearly enough attributed to the use of ADMM (which is not new).

[1] Taming Score-Based Denoisers in ADMM: A Convergent Plug-and-Play Framework, Shrestha et al, ICLR 2026

Questions for authors

see weaknesses

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

This paper studies plug-and-play (PnP) reconstruction with pretrained flow-matching priors. Existing flow-based PnP methods (PnP-Flow, Flower) use a stochastic renoise-denoise step. The authors take its expectation over the renoising noise, and treat it as a deterministic prior operator.

Critical strengths

(1) Writing is easy to follow

(2) Idea is simple and clearly presented.

(3) Proposed algorithm is easy to implement.

Critical weaknesses

The sources of the improvement are not isolated. The method combines four ingredients: flow-based prior, ADMM, per-step averaging, and the late-heavy N_k schedule. Did the authors run controlled experiments that the same ADMM formulation under diffusion prior? Did the authors run controlled experiments testing whether the same averaging helps other denoiser-based solvers? The most relevant are the strongest competitor (Flower with per-step averaging and the late-heavy schedule) and a diffusion-based PnP method such as DiffPIR. Also, one budget-matched table over {PnP-Flow, Flower, ADMM} X {N = 1, constant N, late-heavy}, on the test set with several seeds, would make the contribution of each component clear.

Novelty relative to SNORE, and whether the operator is specific to flows. Up to a change of variables, the mean renoise-denoise operator appears to be the same as the Gaussian-smoothed denoiser used in SNORE (Renaud et al., 2024). In other words, \bar{S}_t is a gradient step on an explicit scalar energy. Sec. 4 claims two distinctions from SNORE: that the operator does not correspond to a scalar energy, and that the construction is designed for the linear interpolation path. This equivalence weakens both, and it implies that nothing in the operator is specific to flow models. Could the authors comment on this equivalence and clarify what is new relative to (annealed) SNORE, beyond replacing stochastic gradient descent with ADMM?

Limited theoretical novelty. Convergence of PnP-ADMM with a deterministic denoiser is well studied, and it is not clear what this analysis adds to existing results. Also, SNORE’s theoretical analysis of their stochastic algorithm also provides convergence analysis. Can author clarify whether replacing the denoised estimate in analysis with average denoised estimate leads to the convergence analsis? Then, can the authors state which parts of the analysis are technically new?

Gap between the theory and the experiments. The guarantees seem to cover little of what is actually run:

  • The analysis concerns the exact mean operator, but the implementation uses N_k = 1 for about half of the iterations.
  • With K = 100, t_k reaches t_max only at the last iteration, so the asymptotic regime of Prop. 2 is never entered. I am curious that if FlowADMM continues past K with t fixed at t_max, does |w_{k+1} − w_k| go to zero, and does PSNR stay stable?

Effect of averaging on perceptual quality. Table 5 shows that PSNR increases steadily with N, but it does not report LPIPS as a function of N. Does a larger N also worsen LPIPS for FlowADMM? If so, could the authors quantify and explain the trade-off, for example by adding LPIPS and SSIM next to PSNR in Table 5?

The intuition in Figure 1 is not verified. Figure 1 argues that \bar{S}_t makes coarse corrections at small t and places the iterate "on the local bump" of the data manifold at large t. The claim is testable: with no data term, FlowADMM reduces to iterating Sbar at t_k, and for an ideal denoiser its fixed points are stationary points of a smoothed log-density (see W1). Could the authors check it on a used flow prior to show how sampled prior looks and with conclusion? Additionally, does this yield improved FID scores in image generation to justify the proposed mean-operator design?

Together with weaknesses above, this leads me to find the novelty and applicability of the paper limited for ICLR in its current form.

Questions for authors

Please refer to the weaknesses section above.

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

This paper presents a new plug & play image restoration algorithm called FlowADMM. The algorithm follows the PnP ADMM algorithm of Ryu et al (2019) with three additional ingredients:

  1. The denoising network DσD_\sigma is replaced by its corresponding mean renoise-denoise operator Sˉσ\bar{S}_\sigma (or a Monte-Carlo estimate thereof), in the spirit of PnP-Flow, Flower and Snore.
  2. The denoiser itself comes from a flow-matching training procedure, but this is conceptually not different from training a denoiser at all possible noise levels σ\sigma.
  3. The ADMM algorithm is annealed, in the sense that the value of σ\sigma is decreased over the iterations until they asymptotically reach a minimum value \sigma_\min>0. The convergence to a fixed point of the proposed method is studied in the case where the data fitting term is strongly convex and the denoiser residual (Dσ−I)(D_\sigma-I) is LσL_\sigma-Lipschitz with Lσ<1L_\sigma <1.

A set of numerical experiments shows that in terms of PSNR, this method performs better than related work in 7 out of 10 tasks, but only in 5 out of 10 tasks in terms of SSIM. For perceptual metrics, however, this drops to 2 in 10 (for LPIPS).

Remark on notation: According to the time schedule t→1t \to 1 means that the noise stdev σt→0\sigma_t \to 0 and conversely.

Critical strengths
  • Good numerical PSNR performance w.r.t. related work
  • Proof of convergence to a fixed point under Lipschitz and strongly convex conditions
Critical weaknesses
  • The novelty of the method is relatively limited. It is essentially an annealed version of Ryu et al (2019)'s PnP-ADMM, with the denoiser replaced by a monte-carlo estimate of a renoise-denoise operator. The convergence analysis also follows the same arguments and limitations as in (Ryu et al 2019), possibly even weaker results (no proof of existence of a fixed point). The only novelty w.r.t. the latter (in terms of convergence theory) is the introduction of annealing instead of a constant σt\sigma_t. And still, in this case, the analysis boils down to a quasi-constant value of σt\sigma_t within a small interval [\sigma_\min, \sigma_\min+\delta] as σt\sigma_t asymptotically tends to \sigma_\min>0. No analysis attempted to explain the benefit of annealing for small values of tt. Like in Ryu et al (2019) the strong convexity condition leaves aside 3 out of 5 problems in the experimental section, namely: super-resolution, random inpainting and box-inpainting are not strongly convex, so they are not covered by the theory.

  • The analysis only shows convergence to a fixed point. There is no discussion on whether this fixed point has any relationship to a critical point of a functional implicit in the limit denoiser D_{\sigma_\min}. However the theory to express an MMSE denoiser as the prox of an implicit functional exists (Gribonval 2011), and has been used in the PnP literature (Xu et al 2020). It is just ignored here. Even worse: in contrast to Ryu's paper, where the denoiser is trained to comply with the Lipschitz condition required for convergence, this work only verifies experimentally whether the denoiser satisfies the condition... and reaches the conclusion that the Lipschitz condition is met in barely more than 50%50\% of the cases (see figure 3).

  • Annealed PnP-ADMM has already been studied in the PnP literature, both for classical denoisers (Chan et al 2017), and Diffusion Models (Kim et al 2026; Shrestha & Fu 2026)

  • The experimental evaluation leaves aside possibly the closest algorithm to FlowADMM, namely the annealed version of SNORE, which is discussed in Marien Renaud's PhD thesis and in Renaud et al's Snore paper that the authors cite.

  • The lack of an interpretation of the implicit denoiser as the proximal operator of a regularization potential R(x) makes it difficult to interpret the consequences of equation (21) in proposition 1. In Ryu's paper this constraint means that there is a limit on how much we can regularize the inverse problem. In practice this leads to suboptimal settings because we have to choose between optimal regularization and convergence. In many cases we cannot have both. How should we interpret equation 21 in this setting.

references

  • Gribonval (2011), Should Penalized Least Squares Regression be Interpreted as MAP Estimation?, IEEE TSP 2011. doi:10.1109/TSP.2011.2107908
  • Gribonval & Nikolova (2020), A Characterization of Proximity Operators, JMIV 2020. doi:10.1007/s10851-020-00951-y
  • Kim, Levy & Wetzstein (2026), Dual Ascent Diffusion, CVPR 2026. arXiv 2505.17353
  • Hurault, Leclaire & Papadakis (2022), Proximal Denoiser for Convergent Plug-and-Play Optimization with Nonconvex Regularization, ICML 2022. arXiv 2201.13256
  • Shrestha & Fu, arXiv 2026. arXiv 2603.10281
  • Xu, Sun, Liu, Wohlberg & Kamilov, Provable Convergence of Plug-and-Play Priors with MMSE Denoisers, IEEE SPL 2020. arXiv 2005.07685 · [doi:10.1109/LSP.2020.3006390]
Questions for authors

Please address the weaknesses signaled above, either by expanding the theory, experiments, or by clarifying any point that may have lead to a misinterpretation on my side.

In addition to these weaknesses, I found several minor issues that you may want to fix or clarify:

  • in line 212 you say that Lemma 1 shows that the "strength of the operator decreases as t→1t\to 1". I could not make sense of this sentence, since you do not define what the "strength" of the operator means. Nor is it obvious from Lemma 1 that the Lipschitz constant of RtR_t decreases as t→1t\to 1
  • Please check in your experiments whether the condition (21) for convergence is satisfied for the chosen value of τ\tau all along the annealed time schedule or not.
  • Appendix 1 in line 256 does not exist. You want to refer to Remark 1 in Appendix A. Please fix.
  • Please add the Snore family of algorithms and particularly their annealed versions to your experimental evaluation
  • In line 292-294 you say that "SNORE relies on additive Gaussian corruption with Gaussian denoisers, while our construction is tailored to the linear interpolation path used by straight-line flow matching." What is the fundamental difference between these two when p0=N(0,Id)p_0 = N(0,Id) is zero-mean white Gaussian with variance=1?
  • In section 5.2 you provide the actual time schedule in equation (25). But you do not check whether this schedule satisfies the condition of Proposition 2, namely equation (22).
  • line 411 you say that your method is optimized for PSNR. Why? A posterior expectation (MMSE estimator) would be PSNR-optimal. But as far as I understand, from a Bayesian perspective, your method computes the MAP estimator (maximum a posteriori) which is not necessarily PSNR-optimal. I mean MAP in the following sense: The data fitting term Fy(x)=−log⁡p(y∣x)F_y(x) = - \log p(y|x) and Gt(x)G_t(x) is the regularization potential implicit in your MMSE denoiser Sˉt\bar{S}_t according to Gribonval (2011). In turn GtG_t is a regularized version of the negative log-prior −log⁡p(x)-\log p(x) . Well your estimator may not be strictly speaking an MMSE denoiser, but may be covered by the more general theory in (Gribonval & Nikolova 2020), to be checked.

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

Related papers

Discussion (0)

Sign in to comment.