acceptodds
Under review as a conference paper at ICLR 2027

Denoising as Projection: Constrained Optimization with Gradient-Guided Diffusion

Abstract

Diffusion models are increasingly used not only for sampling from learned data distributions, but also for generating samples that optimize task-specific objectives. A common approach is to guide the reverse diffusion process using gradients of an external objective. However, when the data distribution is supported on a structured feasible set, such as a manifold or a constraint set, gradient guidance can move samples away from the learned data geometry. In this paper, we study a simple projected-gradient-guided diffusion update based on the observation that the Stein denoising operator can act as an approximate projection onto the data geometry. The proposed update incorporates the objective gradient inside the denoising step, yielding an inference-time method that uses only a pretrained denoiser and gradient evaluations. We analyze this update as an inexact projected-gradient method for constrained optimization over learned feasible geometries. Our theory covers three settings: linear manifolds, compact convex feasible sets, and compact Riemannian submanifolds. In all these settings, we prove descent and finite-time convergence guarantees. Numerical experiments support the theoretical interpretation and illustrate how the proposed update balances objective descent with preservation of the learned geometry.

Primary area:
optimization
Keywords:
diffusion models, constrained optimization, gradient guidance, score-based generative models, projected gradient methods

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_kjMYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes to use a Stein denoiser as an approximate projection operator to map gradient-updated samples back toward the feasible set. It analyzes the approximation error, and the convergence of the method. It considers linear subspaces, compact convex sets, and compact Riemannian manifold.

Critical strengths

The paper provides theoretical analyses of the Stein denoiser as an approximate projection onto the feasible set. The analyses include linear subspaces with Gaussian data, compact convex sets, and compact Riemannian manifold. It also develops convergence analyses for DCG under the stated assumptions.

Critical weaknesses

For the manifold setting, the analysis assumes that the sample is already close to the manifold. It doesn't analyze the case where the sample is far from the manifold. It does not show how the generation process reaches this region from pure noise.

Also the convergence bound only guarantees that there exists a good point along the trajectory, but this does not guarantee that the final generated sample satisfies the same bound.

Questions for authors

Could the authors clarify more about the relationship between DCG and plug-and-play (PnP) proximal gradient methods? Given the extensive literature on PnP convergence, please explain which algorithmic and theoretical contributions go beyond existing results.

2 comments
AuthorsOpenReview ↗

Response to Reviewer kjMY

We thank the reviewer for recognizing our theoretical contributions, particularly the projection-approximation analysis of the Stein denoiser across different feasible geometries. We address the two limitations and the relationship to PnP-PGD below.

1. Local manifold convergence and best-iterate guarantees

(a) Local convergence and initialization from pure noise.

We acknowledge that our manifold convergence guarantee is local. This restriction reflects a fundamental geometric difficulty: for general nonlinear manifolds, the nearest-point projection is only guaranteed to be uniquely defined and smooth within a tubular neighborhood (Federer, 1959, Section 4). Outside this neighborhood, the projection may be nonunique, so the geometric estimates underlying our analysis cannot be directly extended globally. Related methods also assume near-manifold initialization, including score-based optimization on learned manifolds and landing gradient descent on the Stiefel manifold (Kharitenko et al., 2026, Theorems 3 and 5, Ablin et al., 2024, Proposition 9).

Our contribution is to establish convergence within this neighborhood while accounting for both optimization progress and time-varying denoising errors. Establishing convergence from arbitrary Gaussian initialization would additionally require analyzing how the global reverse diffusion dynamics enter this neighborhood, which is beyond our current theoretical scope. Importantly, our experiments initialize DCG directly from Gaussian noise without enforcing this local condition, demonstrating its effectiveness in practice.

(b) Best-iterate versus last-iterate convergence.

We agree that our finite-time nonconvex convergence guarantee applies to the best iterate, rather than necessarily the final generated sample. Such guarantees are standard in nonconvex manifold optimization. They appear in Riemannian gradient and proximal-gradient methods (Boumal et al., 2019, Section 2, Chen et al., 2020, Theorem 5.5) and in methods that allow off-manifold iterates (Schechtman et al., 2023, Theorem 6, Ablin et al., 2024, Proposition 9, Kharitenko et al., 2026, Theorem 5). In general, bounding the sum of stationarity residuals guarantees the existence of a good iterate, but does not imply the same bound for the final iterate without additional assumptions.

Our contribution is to establish this finite-time guarantee for gradient-guided diffusion while simultaneously controlling the Riemannian stationarity residual and distance to the learned manifold under inexact, time-varying denoising. Our experiments evaluate the final generated samples, while extending the theoretical guarantee to the last iterate remains an interesting direction for future work.

2. Relationship between DCG and PnP-PGD

We thank the reviewer for raising this question. We are aware of the PnP-PGD literature and have already discussed its relevance in Section 3 and Appendix A of our original submission. However, classical PnP-PGD does not explicitly incorporate the gradual reverse denoising dynamics of generative diffusion models, and adapting it to gradient-guided diffusion is not straightforward.

Our contribution is to develop a diffusion-compatible gradient-guided optimization algorithm, rather than simply reinterpret existing PnP-PGD methods. In particular, DCG's relaxation coefficient ϵt=1−σt−1σt\epsilon_t=1-\tfrac{\sigma_{t-1}}{\sigma_t} and time-varying Stein denoiser π^t\widehat{\pi}_t are derived from the diffusion noise schedule, rather than introduced as arbitrary algorithmic choices. We further establish the denoiser's approximate projection properties and derive optimization and feasibility guarantees that are not covered by existing PnP-PGD analyses.

We refer the reviewer to our detailed response to Reviewer B4oB (Points 1 and 2) for a more extensive discussion of the algorithmic novelty and theoretical distinctions, including a comparison with related PnP literature.

We hope our response addresses your concerns. If these clarifications resolve your questions and concerns, we would be very grateful if you would consider raising your score.

I thank the author for the detailed response. I acknowledge the technical challenges involved in extending the analysis to nonlinear manifolds from a far point and analyzing the best iterate. The authors have addressed my concerns regarding the relationship to PnP. Therefore, I would like to raise my score to 3.

Reviewer_B4oBOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Denoising-Corrected Gradient guidance, which applies an objective-gradient step before denoising and interprets the resulting update as inexact projected-gradient optimization over learned feasible geometries.

Critical strengths

The method is simple , requiring only one gradient evaluation and one denoiser call per step.

Critical weaknesses

Weaknesses

  1. Limited conceptual and algorithmic novelty. My main concern is the significance of the contribution beyond existing work. Using diffusion models as priors is already well established, and the proposed gradient-then-denoiser update closely resembles a relaxed, time-varying plug-and-play proximal-gradient method. The paper does not sufficiently establish what new capability this formulation provides.

  2. Unclear incremental theoretical contribution. Convergence analysis for plug-and-play proximal-gradient methods has been extensively studied, including under nonconvex formulations (e.g., Hurault et al., 2022). Although the geometry-specific projection bounds are potentially useful, the paper should more clearly distinguish its guarantees and assumptions from existing PnP and inexact projected-gradient analyses. Recovering familiar convergence rates under idealized denoiser assumptions does not, by itself, establish a sufficiently strong theoretical contribution.

  3. Insufficient evidence of an advantage over established methods. The experiments provide some encouraging trajectory-tracking results, but do not establish an advantage over the broader class of diffusion-prior and PnP methods. Direct comparisons with PnP gradient–denoiser baselines using the same pretrained model are particularly relevant. Comparisons on compatible inverse-problem tasks with methods such as DiffPIR and, for linear measurements, DDRM would also help clarify the contribution. The modest D4RL gains and absence of runtime comparisons leave the practical benefits insufficiently demonstrated.

Questions for authors

Could the authors clarify the error condition in Lemma 14? The stated assumption controls ϵEt≤r/4\epsilon E_t\le r/4, whereas the proof on p. 41, lines 2176–2179, uses

76ϵEt≤724ϵr,\frac76\epsilon E_t\le\frac{7}{24}\epsilon r,

which requires Et≤r/4E_t\le r/4. Is an additional restriction on the denoising error needed, and how would this affect Lemma 6 and Theorem 4?

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

This paper studies constrained optimization over a feasible set represented implicitly by a data distribution. The proposed Denoising-Corrected Gradient (DCG) method places an objective-gradient step inside a pretrained posterior-mean denoiser.

The method requires one objective-gradient evaluation and one denoiser evaluation per step, without differentiating through the denoiser. The central interpretation is that the Stein posterior mean approximates projection onto the data support, making DCG a time-varying inexact projected-gradient method.

Critical strengths
  • Optimizing an external objective while preserving structure learned from feasible examples is a useful formulation, particularly when an explicit projection is unavailable. Applying a gradient step before a geometric correction is well motivated.
  • DCG uses only a gradient evaluation and a denoiser call at each step. Avoiding differentiation through the denoiser is a meaningful computational distinction from methods that use denoiser derivatives to construct geometric directions.
Critical weaknesses
    1. Lemma 6's claimed invariance of the half-tubular neighborhood is false under its stated conditions.

The issue is in lines 435–441. The lemma assumes dist⁡(xt,M)≤r/2\operatorname{dist}(x_t,\mathcal M)\le r/2 and a projection error bounded by EtE_t, and claims that the next iterate remains within distance r/2r/2 when the displayed stepsize conditions hold. Controlling ϵEt\epsilon E_t alone is insufficient for this conclusion.

Consider the unit circle M=S1\mathcal M=S^1. Valid chart and curvature parameters are ρ0=rtub=1/2\rho_0=r_{\rm tub}=1/2 and κ=2\kappa=2, giving r=1/12r=1/12. The local graph ϕ(α)=1−α2−1\phi(\alpha)=\sqrt{1-\alpha^2}-1 has ∣ϕ′′(α)∣<2|\phi''(\alpha)|<2 for ∣α∣≤1/2|\alpha|\le1/2, so these choices satisfy the required local regularity. Let f≡0f\equiv0, take the valid upper bounds L=G=1L=G=1, and set η=ϵ=1/100\eta=\epsilon=1/100. Choose

\qquad D_{\sigma_t}(x_t)=\frac{15}{16}e_1, \qquad E_t=\frac1{16}.$$ Then $\pi_{\mathcal M}(x_t)=e_1$ and $\operatorname{dist}(x_t,\mathcal M)=1/24=r/2$. Both stepsize conditions in Lemma 6 hold: $$\eta=0.01\le0.02, \qquad \epsilon=0.01 \le \min\{0.065104\ldots,1/3\}.$$ Nevertheless, the update gives $$x_{t-1}=\frac{1533}{1600}e_1, \qquad dist(x_{t-1}, M) = \frac{67}{1600} > \frac1{24} = \frac r2.$$ This example can be realized by the exact Stein denoiser for the uniform distribution on the circle. Its radial posterior mean is $$D_\sigma(ae_1) = \frac{I_1(a/\sigma^2)}{I_0(a/\sigma^2)}e_1,$$ where $I_0,I_1$ are modified Bessel functions. The ratio continuously ranges from one to zero as $\sigma$ increases and attains the required value at $\sigma\approx0.34023945$. - 2. The description of Guo et al. omits a central geometric component of their method. Lines 207–212 describe the method of [Guo et al. (2024)](https://arxiv.org/html/2404.14743v2) using a bare gradient evaluated at the generated population mean and added outside the denoiser. Their actual look-ahead guidance is $$G_\{ loss\}(x_t,t) = -\beta(t)\nabla_\{x_t\} (y-g^\top D_t(x_t) )^2,$$ which includes a Jacobian-transformed direction $J_{D_t}(x_t)^\top g$ and a residual factor. This geometric transformation is central to their subspace-preservation result; see Definition 1 and Algorithm 1 of the cited paper. The displayed update therefore does not accurately represent that guidance. Table 1 itself lists a Jacobian requirement for Guo et al., making the omission in the text particularly confusing. The comparison should be corrected, and the authors should clarify whether the experimental baseline implements the original method. - 3. The precise novelty beyond existing denoiser-based optimization frameworks is not sufficiently isolated. The algorithm has the form of a relaxed, time-varying plug-and-play proximal-gradient iteration. The paper acknowledges this connection, but the comparison should explain which new assumptions, approximation bounds, or convergence arguments make the contribution substantial. This is especially relevant to [Leong and Traonmilin (AISTATS 2026)](https://proceedings.mlr.press/v300/leong26a.html), who already interpret noise-dependent diffusion denoisers as approximate projections and analyze generalized projected-gradient iterations for inverse recovery.
Questions for authors
  1. Which projection bounds and convergence arguments are new relative to existing PnP theory and Leong–Traonmilin's time-varying projection framework? What theoretical or empirical benefit does the DDIM relaxation provide over direct PnP-PGD with the same denoiser?

  2. For manifold experiments, how is entry into the required local neighborhood achieved, and how frequently do guided trajectories fail to enter or remain in the specified neighborhood?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Fwzy (Part I)

We sincerely thank the reviewer for the careful and constructive review! Below, we address the reviewers questions and concerns point by point. We would also like to note that all the discussions are integrated to the revised version of the paper (highlighted in blue)

1. Half-tubular neighborhood invariance in Lemma 6

Thank you for reading the paper so carefully and for providing a clear counterexample! It helped us identify a small gap in the tubular-invariance proof. The original condition was ϵEt≤r4\epsilon E_t\le \tfrac{r}{4}, but the proof requires Et≤r4E_t\le \tfrac{r}{4}, where EtE_t bounds the denoiser’s projection error. Your example satisfies the original condition, with ϵEt=11600\epsilon E_t=\tfrac{1}{1600}, but is excluded by the corrected condition, since Et=116>148=r4E_t=\tfrac{1}{16}>\tfrac{1}{48} =\tfrac{r}{4}.

The correction is local to the tube-invariance step in Lemmas 6 and 14. The normal-distance recursion uses 76ϵEt≤724ϵr\tfrac76\epsilon E_t\le\tfrac7{24}\epsilon r, which follows from Et≤r4E_t\le \tfrac{r}{4}. Together with the existing curvature bound, this ensures that an iterate within r2\tfrac{r}{2} of the manifold stays within r2\tfrac{r}{2} after the update.

The condition is mild in the low-noise regime of Theorem 4. The radius r>0r>0 is fixed by the manifold geometry, and the Stein projection error bound Cσt∣log⁡σt∣C\sigma_t|\log\sigma_t| tends to zero as σt→0\sigma_t\to0. We choose the starting noise level so that this bound is at most r4\tfrac{r}{4} and σt0≤e−1\sigma_{t_0}\le e^{-1}. It then stays below r4\tfrac{r}{4} as the noise decreases. This closes the backward induction and yields the stated Lyapunov descent inequality and best-iterate convergence rate on the low-noise segment.

2. Correction to our description of Guo et al.’s algorithm

Thank you for carefully checking our comparison with Guo et al. We agree that our original description did not accurately express their look-ahead guidance. We acknowledge this writing error and have corrected it in the revised manuscript. We would also like to clarify that our GGDM implementation follows the original algorithm, including the batch-based outer loop and the full look-ahead guidance during reverse sampling.

The revised passage reads as follows.

Guo et al. evaluate gk=−∇f(xˉk)g_k=-\nabla f(\bar x^k) in our minimization convention, where xˉk=B−1∑b=1Bx0k,b\bar x^k=B^{-1}\sum_{b=1}^{B}x_0^{k,b} is the mean of the BB terminal samples from outer iteration kk. They hold gkg_k fixed while generating the next batch and add the look-ahead guidance

G_k(x_t,t) =-\beta(t)\nabla_{x_t}\bigl(y_k-g_k^\top\widehat\pi_t(x_t)\bigr)^2 =2\beta(t)\bigl(y_k-g_k^\top\widehat\pi_t(x_t)\bigr) D\widehat\pi_t(x_t)^\top g_k \end{aligned}

to the score. Here, yky_k is the target for the linearized reward and β(t)\beta(t) is the guidance weight. Both the residual and the Jacobian-transformed direction depend on the current noisy state.

AuthorsOpenReview ↗

Response to Reviewer Fwzy (Part II)

3. Contribution relative to PnP-PGD and Leong–Traonmilin

Thank you for asking us to clarify the contribution. Here, we summarize the key contributions and refer the reviewer to our response to Reviewer B4oB, who raised a similar concern, for a more detailed discussion. Our central conceptual contribution is a diffusion-compatible gradient-guided optimization algorithm (see response to Reviewer B4oB, Point 1). Classical PnP-PGD does not explicitly incorporate the gradual reverse dynamics of generative diffusion models, and adapting its gradient–denoiser composition to these dynamics is nontrivial. In DCG, the diffusion schedule determines both the denoising level and the relaxation coefficient ϵt=1−σt−1σt\epsilon_t=1-\tfrac{\sigma_{t-1}}{\sigma_t}. Deriving and justifying this formulation is itself a contribution. The theoretical benefit (see response to Reviewer B4oB, Point 2, and the comparison with Leong–Traonmilin below) is to establish constrained-optimization guarantees beyond the fixed-denoiser, regularized-objective setting of the cited PnP analysis. We cover linear-Gaussian subspaces, compact convex supports with general strongly convex or smooth nonconvex objectives, and local optimization on smooth manifolds. The analysis is also more challenging, because the relaxed reverse trajectory requires jointly controlling objective descent, residual infeasibility, and denoising errors that change with the diffusion schedule. The empirical benefit of DCG is clearest on complex, higher-dimensional, nonconvex trajectory optimization tasks (see response to Reviewer B4oB, Point 3). In comparisons using the same pretrained denoisers, fixed-noise PnP-PGD can outperform DCG on simple convex tasks, while DCG generally achieves better tracking during execution, greater dynamic feasibility, and higher D4RL returns on the trajectory benchmarks.

Regarding Leong–Traonmilin, their algorithm uses diffusion denoisers, but its update is not a DDIM reverse step. Their iteration is xn+1=Dσn(xn)−μA⊤(ADσn(xn)−y)x_{n+1}=D_{\sigma_n}(x_n)-\mu A^\top(AD_{\sigma_n}(x_n)-y), which applies denoising followed by a measurement-gradient correction. Although the denoising level decreases, this update does not interpolate between the current noisy iterate and the denoised prediction using consecutive noise levels. DCG preserves this DDIM interpolation, with weights σt−1σt\tfrac{\sigma_{t-1}}{\sigma_t} and 1−σt−1σt1-\tfrac{\sigma_{t-1}}{\sigma_t}, while incorporating objective guidance. Further, their nonconvexity concerns the feasible set, while their objective remains the convex quadratic measurement loss 12∣∣y−Ax∣∣2\tfrac12 ||y-Ax||^2. Specifically, their nonconvex recovery result treats unions of subspaces induced by low-rank Gaussian mixtures under restricted-isometry assumptions. Our analysis accommodates general strongly convex and smooth nonconvex objectives and develops local guarantees for Riemannian stationarity and feasibility on curved smooth manifolds. We have added the above discussions into the revised version of the paper in Section 5.2 and Appendix A.

4. Entry into and retention within the manifold neighborhood

We thank the reviewer for this question. Our experiments initialize DCG directly from Gaussian noise, without explicitly enforcing entry into or retention within the tubular neighborhood required by our local convergence analysis. Empirically, the gradual reverse denoising process brings samples toward the learned data manifold. Our experiments also evaluate dynamical feasibility, which provides an empirical measure of how closely the generated trajectories satisfy the underlying system dynamics. The consistently strong feasibility results suggest that DCG effectively approaches and remains close to the feasible geometry in practice.

I thank the authors for their detailed responses. My concerns have all been addressed. I would like to keep my current score rather than raise it, since — as reflected in my confidence score above — I am not very familiar with this field.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Reminder and summary of revisions

Dear Reviewers,

Thank you once again for your review comments! As the discussion period draws to a close, we wanted to send a brief reminder that our detailed response and revised manuscript are available in this discussion. For ease of reading, we’ve summarized the main changes below.

We ran new baseline comparison with PnP-PGD using the same pretrained denoisers. The baseline had higher tracking error during execution on all three unicycle references and lower returns on all three D4RL tasks. We also expanded our comparisons with PnP-PGD (Hurault et al., 2022), DiffPIR, DDRM, and Leong and Traonmilin (2026), whose work also interprets time-varying denoisers as approximate projections for linear inverse recovery. We clarified how DCG incorporates gradient guidance into the DDIM reverse step, with the diffusion schedule determining both the denoising level and the interpolation weight. Finally, we corrected the denoising-error condition to Et≤r/4E_t \le r/4 and updated the manifold convergence proof.

If you have any remaining questions or concerns, we’d appreciate hearing from you soon so we have time to respond before the discussion closes on November 18. If the revisions address the points you raised, we would be really grateful if you would consider updating your overall assessment. Thank you again for your time and feedback!

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.