- Rating
- 1
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
Paper propose a MAP estimation for imaging inverse problems, based on noise annealing and averaging of MMSE denoiser (Gamma). The algorithm consists of two step (1) gradient step on adjusted data-fidelity $ (\lambda(\sigma^2)/2)||x − u||^²$, which adds a vanishing Tikhonov anchor toward a chosen reference point u (2) denoising the current step.
Three variants is proposed including deterministic Gamma (paper proves that the iterates are bounded and every cluster point is a MAP solution. If additionally , the iterates converge to , the MAP solution closest to u.), renoised-GAMMA (the same result when the denoiser output is averaged over infinitely many noisy copies of the input.), stochastic-GAMMA (convergence in mean square for a single-sample version, but only if the injected noise is decoupled and much smaller than the denoiser level).
Experiments use annealed models on CelebA-128 and AFHQ-Cat-256 for denoising, deblurring, ×4 super-resolution, random inpainting and box inpainting. Baselines are PnP-Flow and Approx-PGD.
Critical strengths
- A single-loop algorithm that extends MMSE Averaging (Pesme 2025) from MAP denoising to general inverse problems.
- A unifying view where with λ = 0, GAMMA is annealed RED. Renoised-GAMMA with one sample is annealed SNORE.
- Convergence of the iterates themselves to a specific MAP point.
Critical weaknesses
- Novelty is thin. Some claims and prior work question the novelty of the work.
- “First approach to bridge provably convergent denoiser methods and annealed generative methods”
- Shrestha & Fu, ICLR 2026: Score-based denoiser with ADMM, noise added to the denoiser and prove two convergence results (1) high-probability fixed-point-ball convergence with a constant step, and, (2) under weaker conditions, convergence with an adaptive step-size schedule via a bounded-denoiser argument.
- Chan IEEE TCI 2017: Their main theorem proves fixed-point convergence of PnP-ADMM for any “bounded denoiser”, where the continuation scheme drives the penalty to infinity, which results in annealing the denoiser strength.
- Park SGPnP: Similar algorithm with data-fidelity gradient + noise injected to the MMSE denoiser. Their regularizer is exactly the prior term of GAMMA’s renoised objective. Injected noise level differs from the conditional noise level denoiser is trained on like Gamma stochastic. SGPnP claims that under an annealed noise schedule its iterates converge to a stationary point of the exact, un-smoothed objective, (MAP objective).
- The vanishing Tikhonov anchor and selection of the MAP point closest to u is from classical convex optimization (Attouch & Cominetti (1996), Cominetti, Peypouquet & Sorin (2008)).
-
Assumptions: Authors justification on assumption H2 is not satisfactory. Log-concave is unimodal, and this matters here specifically, noiseless Gamma and Approx-PGD objective is MAP, where we see the cartoon like images , here rejoiced and relaxed version work fine. This shows that good results are consistent with not converging to MAP. The cited papers that use this assumption are Langevin sampling, very different setting.
-
Exact MMSE denoiser might create an issue in the assumed setting, while being fine in the other algorithms, because the contraction per step is only , there is nothing to offset the effect of inexact MMSE denoiser.
-
The proof contracts the distance to the moving minimizer by a factor per step. Since , we have . So the total guaranteed contraction after 2000 iterations is at best . The initial error shrinks by at most about 85%. The is not consistent with the theorem.
-
The selection condition is far from met. With : , so and . That ratio is supposed to go to zero, yet it is about 60. Even with σ₀ = 1 it is about 0.6. So the “converges to the MAP closest to u” result says nothing about these runs. The paper also never states what u is in the experiments.
-
Experimental resutls:
- The method aimed at the MAP is worse than the degraded input on deblurring (both AFHQ and Celeba). Deblurring is overall not a difficult inverse problem, around 10 dB worse seems to be something gone wrong like iterates not converging, denoiser far from exact MMSE, MAP is bad estimator.
- Could the authors report the convergence of iterates, like residual over iteration?
- Performance is bad, compared to PnP-Flow.
- Compute is not matched, this method seems to be comparably more expensive.
- Baselines are not included: Generative: only PnP-Flow is included. Missing are DiffPIR, DPS, RED-diff, and Flower, ICTM, D-Flow, SGPnP and annealed SNORE. Convergent methods: the abstract claims GAMMA “surpasses convergent methods,” but the only convergent baseline is Approx-PGD, which appears broken (point 1). The standard convergent PnP methods, GS-PnP and Prox-PnP (Hurault2022), are not compared.
- No evidence of convergence on real images
Questions for authors
how does an approximation error in the denoiser propagate through the analysis, and what decay of preserves convergence?
Are the schedule assumption same as reported runs? It says , however, .
What is u in the experiments, and does changing it affect results?
3 comments
We now address the remaining remarks and questions of the reviewer.
Discussion on the assumptions (H2)
"Authors justification on assumption H2 is not satisfactory."
We have edited the discussion below H2 (Section 4.1) to emphasize that, while log-concavity is a strong requirement, our proof is inherently local, which suggests that an extension to multimodal, locally log-concave distributions is possible. This perspective is supported by our experimental results on a Gaussian mixture prior and on natural images.
"noiseless Gamma and Approx-PGD objective is MAP, where we see the cartoon like images , here renoised and relaxed version work fine. This shows that good results are consistent with not converging to MAP. "
The artifacts of noiseless GAMMA and Approx-PGD do not come from converging to the MAP, and the good results of Renoised and Relaxed GAMMA do not come from diverging away from the MAP:
- As now addressed in Section 5.1 (Figure 1, 2, 10, 11), when the MMSE denoiser is exact both noiseless GAMMA and Renoised-GAMMA converge to the same MAP solution, whereas PnP-Flow and DPS fail to converge.
- We believe that the performance gap between GAMMA and Renoised/ Relaxed GAMMA on the natural images comes from the denoiser rather than from the MAP estimator itself. This diagnosis is consistent with Vert et al. (2026, already cited in Section 5.3), who show that the cartoon-like artifacts of algorithms targetting the MAP with Diffusion Models "seems to come from score errors" rather than from MAP estimation itself. Indeed, a neural denoiser being trained on noisy images, its evaluation on unnoised iterates is inaccurate.
Bad result for deblurring
"Deblurring is overall not a difficult inverse problem, around 10 dB worse seems to be something gone wrong"
Typo on our part in the Table 4: the super-resolution and deblurring entries of the GAMMA (noiseless) row were accidentally swapped, creating this 10dB gap. It is now fixed.
Cost of the method, guarantees for finite schedules
The slow convergence of GAMMA and Renoised-GAMMA is caused by the conservative choice of the schedule, not the method itself. Relaxed-GAMMA (Renoised-GAMMA with more aggressive schedules) converges in 100-500 iterations while producing high quality reconstructions for every inverse problem. We added Corollary 9 (Section 4.3) which establishes that, for a family of finite schedules, the final iterate converges (at a given rate) to the MAP, when grows.
Additional baselines (DPS and DiffPIR)
"Baselines are not included"
We have added DPS and DiffPIR to our benchmarks on both CelebA (Table 1) and AFHQ-Cat (Table 4). Relaxed-GAMMA consistently outperforms DPS and DiffPIR while matching PnP-Flow.
"The standard convergent PnP methods, GS-PnP and Prox-PnP (Hurault2022), are not compared."
In our comparison, we restrict ourselves to methods relying on off-the-shelf denoisers, which excludes Gradient-Step (GS) denoisers. We have clarified the sentence in the abstract.
Evidence of convergence on images / choice of / contraction factor close to 1
We have added a new paragraph in Appendix D.2 with , so that is times larger than in the main experiments. This regularization strength increases the impact of , makes go to 1 less fast, and go to 0 faster (e.g. contraction factor goes from to ). We also run k iterations instead of k, to get as close as possible to the asymptotic regime. For those examples, we added in Figures 7, 8 and 9 curves for the residual along the iterations , together with the MAP objective (where is numerically estimated via the learned denoiser).
We observe the following:
- Impact of u: in Figures 7, 8 and 9, an informative choice for (e.g. an average dataset image for inpainting that fills the mask, bicubic interpolation for superresolution) strongly improves the reconstruction, especially for the noiseless version of GAMMA. On superresolution in particular, we avoid the artifacts that originally appeared in the main experiments and achieve a PSNR far above the observation.
- On convergence on images, both GAMMA and renoised GAMMA decrease the MAP objective and have low residual along the iterations. The convergence remains slow in these experiments, but the schedules used in the Relaxed GAMMA version suggest that accelerated convergence is achievable.
- Renoised GAMMA reaches substantially lower MAP objective values than its noiseless version at equal iteration count, which is consistent with its better reconstruction metrics in our experiments.
Questions
Handling approximate MMSE
We added results in Appendix B.5. Similarly to what is shown in Pesme et al. 2025, when the datafit term is strongly convex and the score error over the iterates is bounded by , the iterates converge to a point at distance O(ξ) of the MAP. We believe a further study would deserve a separate work.
Are the schedule assumption same as reported runs
The schedules used in the runs are defined in Remark 8 and their parameters are indicated in Table 2&3. For GAMMA and Renoised-GAMMA, every schedule verifies the conditions of the theorems.
What is and does changing it affect the result ?
In the main experiments (Tables 1 and 4), we set (now explicitly stated at the beginning of Appendix D, l. 1682), which has a little impact on PSNR because is small. In contrast, when is larger (, Appendix D.2, Figures 7–9), choosing an informative (dataset mean image or bicubic interpolation) significantly improves PSNR and removes artifacts, as detailed above.
We thank the reviewer for their thorough reading and constructive feedback, and recognition of our paper's strengths (single-loop extension of MMSE Averaging to general inverse problems, unifying view of annealed RED/SNORE, and iterate convergence to a specific MAP estimator). Below, we first address the concern about the limited novelty of our method.
Novelty
Thank you for pointing out these interesting references, which we have carefully read. However, in our view, they do not diminish the novelty of our work.
"First approach to bridge...” (l.479): In Section 6, we wrote "convergent denoiser-based MAP methods, such as Pesme et al. (2025)", referring to convergence to a minimizer of the true unsmoothed objective F. Our goal is to characterize the limit rather than showing that the iterates converge to an unknown limit.
Chan et al. (IEEE TCI 2017) & Shrestha & Fu (ICLR 2026) : The first paper is already cited in our Introduction (l.037–042), where we added the second one. In the two articles, it is proven that the iterates form a Cauchy sequence and thus converge. However, when using denoisers with noise levels going to 0, the limit point is not characterized and is not shown to minimize any underlying objective.
Vanishing Tikhonov (Attouch & Cominetti, 1996; Cominetti et al., 2008) As acknowledged in Section 3 ("Related works", l. 243–245, citing Attouch & László, 2024), vanishing Tikhonov is a classical convex optimization technique. We will add these pioneering references. Our contribution lies in coupling it with the non-standard Gaussian smoothing inside the logarithm (−logpσ) to make the algorithm work when the data-fidelity is not strongly convex, in the context of inverse problems.
Park et al. (SGPnP, 2026) : Thank you for pointing out this recent work. While SGPnP is indeed close to a single sample Renoised GAMMA without Tikhonov regularization, there are fundamental distinctions in what is proven, as well as unaddressed theoretical limitations and incompatible assumptions in its analysis, that our work overcomes.
Main distinctions:
The annealed convergence of SGPnP (Theorem 2) is proven (also in the setting where the denoisers are perfect MMSE denoisers) with a double-loop regime requiring Ik inner iterations for a fixed σk. By their Assumption 3 and Lemma 3, one can show that Vmax(σ)≥1/σ2, so that Ik→∞ as σk→0. Moreover, no explicit step size schedule is given beyond the constraints γk≤1/Lk and γkLkVmax(σk)→0. The proof of Theorem 2 assumes iterate compactness (Assumption 4). In contrast, for (Renoised-)GAMMA, we prove convergence of the single-loop iterates (Ik=1) to a uniquely characterized point x†=ProjArgmin F(u), and we provide a family of explicit schedules αk=σk2+τσk2. We do not assume anything on the iterates.
Theoretical limitations and incompatible assumptions: Several assumptions and proof steps are mathematically problematic.
Assumption 1 requires ∇fσ=∇g+∇hσ to be uniformely bounded on Rd, which fails for quadratic data-fidelity terms g(x)∝∥Ax−y∥2. Assumption 2 vs Assumption 3: By Assumption 2 and Tweedie's formula, Uσ(x,n)=−σn−∇logpσ(x+σn). Requiring ∥Uσ(x,n)∥≤B almost surely (Assumption 3) is therefore impossible for n∼N(0,Id). Assumption 4 and proof of Theorem 2: Assuming that the stochastic iterates remain within a compact set K (Assumption 4) is strong, and its justification in the first lines of the proof of Theorem 2 is invalid. Furthermore, the uniform bound Δ<+∞ on fσk(xk(0))−fσk⋆ is unproved, and the use of Bolzano-Weierstrass in the end is improper. They do not show the convergence of the iterates, but rather that there exists a subsequence of iterates for which ∇f0 goes to 0 almost surely.