acceptodds
Under review as a conference paper at ICLR 2027

Primal-Dual Constrained Diffusion Sampling

Abstract

Many applications call for generative models to provide samples that satisfy requirements, including target semantic or visual features in image generation or observation consistency for inverse problems. Although diffusion models excel at capturing complex distributions, their stochastic, iterative nature makes it difficult to impose such requirements. Typical approaches, known as guidance, tend to encourage rather than enforce desired properties and are often sensitive to hyperparameters such as guidance weights and step size schedules. Moreover, they typically incur substantial memory and computation costs due to retraining or backpropagating through the diffusion model, which limits the type and number of requirements they can impose. This work addresses these challenges by casting constrained diffusion sampling as an optimization problem in which each requirement is described as generalized moment constraint. Using duality, it shows that the solution to this problem is achieved by a reverse dynamic akin to a parametrized Doob’s h-transform of the original diffusion process, whose parameters are the problem’s Lagrange multipliers. It then introduces Primal-Dual Constrained Diffusion Sampling (PD-CDS), a training-free algorithm that automatically adjusts these parameters throughout the diffusion, and derives convergence guarantees. The generality of this approach is illustrated by inverse problems, semantic control tasks, and visual property constraints, separately and combined. By precluding backpropagation through the diffusion model, PD-CDS is able to match and even outperform existing methods while reducing computation and memory costs.

Primary area:
generative models
Keywords:
Diffusion models, Guidance, Constrained optimization, Duality

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_29qoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes PD-CDS, a guided sampling method for diffusion models that allows for the incorporation of various constraints during the sampling process. Like many other approaches, PD-CDS is training-free; it applies guidance to the posterior mean estimate, x^0∣t\hat{x}_{0|t}, during diffusion sampling. A key distinction of this method is that the step size, λt\lambda_t, is derived from a dual problem and updated automatically. In our experiments—which utilized a simple unconditional image generation diffusion model—we addressed basic inverse problems in imaging as well as the generation of human face images that satisfy specific, user-defined conditions. In the latter task, PD-CDS demonstrated superior performance compared to other existing methods.

Critical strengths

In the task of generating samples of human faces that satisfy two specific criteria, this approach demonstrated strong performance compared to other methods. This suggests that similar techniques could potentially be applied in other domains as well.

Critical weaknesses

Weakness 1: Although the authors claim to have automated the hyperparameter λ\lambda, the method still requires ηt\eta_t to determine λt\lambda_t. Furthermore, hyperparameter tuning can effectively be performed indirectly by varying the definition of gig_i. Providing an ablation study—specifically regarding ηt\eta_t or the different definitions of gig_i—to demonstrate the algorithm's robustness would significantly strengthen the case for its utility.

Weakness 2: In Table 1, it appears that TV regularization could potentially be applied to other algorithms as well; however, no such comparison was conducted. Additionally, it would be beneficial to demonstrate the utility of the proposed method across a broader range of domains, beyond relatively simple experiments like those using the CelebA-HQ or FFHQ datasets.

Questions for authors
  1. How did you choose ηt\eta_t? I wonder whether ηt\eta_t is a sensitive hyperparameter.

  2. What would happen if you define gig_i in other ways? For example, l2-norm-square vs l2-norm

5 comments

We thank the reviewer for their insightful comments. We address each of their concerns point by point below. Additionally, we are preparing a revised version of the manuscript including these new experiments and comments (we thought it preferable to give priority to the responses). We will let them know as soon as the revised version is ready.

Question 1/ Weakness 1

The step size η\eta is indeed a hyperparameter of PD-CDS, as we note in the manuscript (line 232), but it remains a scalar hyperparameter regardless of the number of constraints or diffusion steps (e.g., the 4 constraints in Fig. 1). This is a key advantage of PD-CDS: by tuning a single step size, we obtain a sample-specific, dynamic guidance schedule. This is in contrast to conventional guidance methods that require tuning the relative weight of each constraint jointly with their method-specific hyperparameters, a procedure that is challenging even for a moderate number of constraints and that does not necessarily yield better results (see answer to Question 1.1 to reviewer aS9Q).

In practice, we observe that PD-CDS is not particularly sensitive to η\eta. As the reviewer suggested, we conducted an ablation study of η\eta on the Smile + brightness experiments. The results are summarized below (performance is evaluated on 50 images):

Dual LRJoint AccuracyKID
0.02548%0.04121
0.05082%0.04267
0.10096%0.04210
0.20098%0.03860
0.350100%0.03348
0.550100%0.02376
0.750100%0.08034

These preliminary results indicate that a broad range of η\eta achieves competitive results. In fact, PD-CDS outperforms the baselines for almost all step size values. We observe a similar robustness in the experiments suggested by reviewer Zaeu. There, we change the number of diffusion steps and obtain similar performance using the same η\eta (see response to Weakness 2). We will include these results in the revised manuscript to showcase the robustness of PD-CDS.

Question 2 / Weakness 1

We believe there is a fundamental difference between gig_i and η\eta. Indeed, the constraint functions gig_i are part of the problem definition, namely, part of the optimization problem (PI) (line 143). The step size η\eta, on the other hand, is a hyperparameter of Algorithm 1 which targets (PI). Thus, changing gig_i changes the optimization problem itself, including its feasibility set and geometry, therefore affecting any algorithm used to solve it. In our view, this is fundamentally different from tuning an algorithmic hyperparameter such as η\eta.

That being said, we agree that the choice of gig_i can also affect the algorithm. As requested, we have therefore conducted new experiments on nonlinear and motion deblurring with different gig_i. The results are summarized below.

TaskConstraint functionFIDLPIPSPSNR (dB)Avg. relative violation
Motion deblurringL145.800.06229.370.295
L2^2108.200.31319.400.851
Nonlinear deblurringL133.200.12633.301.05
L2^233.200.04830.920.075

As expected, changing gig_i changes the problem and therefore the performance of the solution. Nevertheless, all the constraints are largely satisfied (relative violation less than 1), except for nonlinear deblurring. Yet, note that the constraint tolerance is set to 0.01, so that a 5% violation corresponds to an excess of 0.0005 on the original intensity scale. Since the expectation is estimated over 100 images, we cannot rule out that the final distribution is feasible. Interestingly, the ℓ1\ell_1- and the ℓ2\ell_2-norm formulations do not consistently outperform one another (ℓ1\ell_1 is better for motion and ℓ2\ell_2 for nonlinear deblurring). This is an important advantage of PD-CDS, as it can accommodate a wide variety of gig_i (and their combinations).

Weakness 2a

Indeed, TV regularization could also be added to other methods. Yet, it is important to highlight that this is not as straightforward and natural as in the case of PD-CDS. To underscore this point, we conducted experiments incorporating TV guidance into DPS and RED-Diff on motion and nonlinear deblurring (see table below). Notice that while TV regularization improves the performance of DPS on both tasks, its effect on RED-Diff is less regular. This is to be expected given that these methods are not designed to naturally accommodate multiple constraints. Overall, PD-CDS achieves better metrics on both tasks.

TaskMethodFIDLPIPS
Motion deblurringPD-CDS56.410.075
PD-CDS + TV46.120.061
DPS85.670.169
DPS + TV71.940.158
RED-Diff63.940.117
RED-Diff + TV60.040.109
Nonlinear deblurringPD-CDS35.380.150
PD-CDS + TV29.650.107
DPS284.710.626
DPS + TV276.100.605
RED-Diff110.640.339
RED-Diff + TV130.410.346

Weakness 2b

We acknowledge that evaluating PD-CDS in different domains would better showcase its utility, but disagree that the experiments provided are “relatively simple.” Image generation and restoration are standard benchmarks for evaluating diffusion-based (guidance) methods and are used across our baselines. Our experiments also go beyond typical settings by simultaneously enforcing up to four heterogeneous constraints, a significant additional challenge compared to single-constraint tasks. Finally, note that PD-CDS is being applied to two tasks that are often tackled by different algorithms (conditional generation and inverse problems), as evidenced by the different sets of baselines we deployed. Nevertheless, if the reviewer had a specific additional experiment in mind, we would gladly consider it (within the rebuttal time constraints).

I appreciate the authors for the thoughtful response.

The ablation results in Table 10 show that several step sizes give similar performance in one setting. However, I still think this alone does not show that choosing the step size is easy or that the method is generally insensitive to it.

For Weakness 2a, the authors report experiments on motion deblurring and nonlinear deblurring. These are only two of the five settings in Table 1 (which is not updated in the revision), and they are also the settings where PD-CDS already performs best. This makes the comparison feel selective, rather than a fair check of how TV regularization affects the other methods. It would be more convincing to add TV regularization to the other methods in the remaining three settings and update Table 1 with the results.

For Weakness 2b, I thought like DIV2K was a natural choice. Many image restoration works use DIV2K. It would be much more great if the authors could do that cause, if so, they could insist generalizability.

For those reasons (especially concerns about generalizability), I would like to maintain my score.

P.S.) I think the abbreviation for 'assumption' is 'asm,' which looks better than 'ass.'

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

This paper casts training-free guided diffusion as a constrained KL-optimization problem and solves it in the dual: the constrained target is shown to be reachable via a parametrized Doob-like h-transform of the pretrained reverse process, whose parameters are the Lagrange multipliers. The authors then propose PD-CDS, a training-free sampler that alternates primal diffusion steps with projected dual updates on a single trajectory, without backpropagating through the model. A finite-horizon convergence guarantee is provided, and experiments on inverse problems and multi-attribute image control show competitive quality, roughly halved memory, and effective handling of multiple heterogeneous constraints.

Critical strengths
  • Reframing guided diffusion as constrained optimization with a duality-based Doob hh-transform is a clean, novel conceptual step.

  • The duality result, the h∗h*-transform characterization, and the finite-horizon dual convergence bound form a coherent and non-trivial chain, with careful error decomposition.

  • PD-CDS avoids backpropagation through the DDM and requires no per-constraint guidance-scale tuning, yielding roughly half the peak memory of DPS/TFG/UGD/LGD/FreeDoM and competitive wall-clock time on nonlinear tasks

Critical weaknesses
  • The analysis in Thm. 4.4 is carried out for an idealized scheme under Ass. 4.3, but Alg. 1 relies on several approximations that are not reflected in the stated error terms. As a result, the guarantee characterizes an idealized sampler rather than PD-CDS itself, making the paper's convergence claim weaker than presented. The authors should either bring these approximation errors into the bias bound or explicitly scope the guarantee to the idealized setting.
  • The efficiency claim would be more convincing with a quality-vs-NFE comparison. As the NFE counts in Tables 9–10 show, PD-CDS operates at a substantially higher NFE budget than several baselines; a plot of FID/KID against NFE (or a matched-NFE comparison) would clarify whether the reported gains persist at comparable compute
Questions for authors

see the weaknesses.

1 comment

We thank the reviewer for their insightful comments. We address their concerns point by point below. We are also preparing a revised version of the manuscript including new experiments and their comments. We however thought it preferable to give priority to our responses. We will let them know as soon as the revised version is ready.

Weakness 1

Though our analysis does rely on certain approximations (Assumptions 4.2-4.3), we would like to clarify that Theorem 4.4 analyzes Algorithm 1 and not the idealized scheme in (8). The only “idealization” (error that we do not account for) in our results concerns the pretrained diffusion model, that we assume to be the exact reference score (Assumption 4.3). As we mention in the manuscript (line 354), we do so to isolate the approximation errors in PD-CDS from those already existing in the underlying DDPM and already studied in, e.g., (Chen et al., 2023; Silveri & Ocello, 2025; Conforti et al., 2025). A combined analysis is of great interest, but beyond the scope of this paper.

Aside from the score, our analysis explicitly accounts for all approximations introduced by PD-CDS. We agree that their impact on the bias could have been made clearer in Thm. 4.4, but due to space constraints we simplified the residual bias to C=O(R(1+εh+εGD))C = O(R(1+\varepsilon_h+\varepsilon_{GD})) and relegated the explicit form of (14) to Thm D.1 (Appendix D). Comparing (14) to (18), we notice that the O(R)O(R) term in (14) actually incorporates the approximations introduced in Section 4, namely,

  • Posterior/Tweedie approximation (eqs. 9-10), i.e., the error of evaluating the constraint at the conditional mean instead of computing its conditional expectation; and
  • Readout approximation (eqs. 10-11), i.e., the error of using an approximate guided score in Tweedie's formula.

Both errors are explicitly bounded in Appendix D.3.

Finally, the errors incurred from updating the multipliers online, along a single diffusion trajectory, are explicitly controlled in Appendices D.4 to D.6. A key contribution of our results is in fact showing that the additional contributions arising from dual variable updates vanish as the number of diffusion steps TT grows. Indeed, they decrease with O(T−1)O(T^{-1}) and O(T−1/2)O(T^{-1/2}) (see Thm D.1).

Weakness 2

The reviewer has a point, although we believe we may have caused some confusion with our terminology. Indeed, by “NFE” we mean the “number of diffusion steps,” which for guidance methods may have substantially different computational and memory costs. We will clarify this ambiguity in the revised manuscript.

That being said, note from Table 9 that PD-CDS uses the same number of diffusion steps (1000) as most competitive baselines (namely, DAPS, DPS, and RED-Diff). Yet, though it uses the same number of steps as those baselines, PD-CDS is faster and more memory efficient. Only DiffPIR and PiGDM are considerably faster, using only 100 steps. Nevertheless, DiffPIR performs poorly on our nonlinear tasks while PiGDM cannot handle them in our experimental setup. Table 10, on the other hand, normalizes the sampling time per image (a surrogate for compute) rather than the number of diffusion steps since the methods have substantially different computational costs per step (PD-CDS being roughly 2.3 faster per step than the baselines).

Still, as the reviewer suggested, we conducted an additional experiment on the High Contrast + Not Young task using 100 diffusion steps, matching the number of diffusion steps for the baselines in Table 8. Importantly, we use the same learning rate as before (see results below). Even at 100 diffusion steps, PD-CDS achieves better KID and joint accuracy than all other baselines. In fact, the results almost match those obtained with 250 steps.

MethodStepsJoint Acc.KIDTime/imagePeak VRAM
PD-CDS10084.5%0.05195.93 s1,298.0 MiB
PD-CDS25086.0%0.044415.14 s1,298.0 MiB
TFG10074.5%0.058612.23 s2,605.0 MiB
UGD10023.5%0.081817.07 s2,607.0 MiB
LGD10016.5%0.105016.42 s2,607.0 MiB
FreeDoM10070.5%0.062016.33 s2,607.0 MiB

Producing results at a lower number of diffusion steps for all baselines would be infeasible in the time of this rebuttal, but we hope that these preliminary results are sufficient to convince the reviewer that they would not change our conclusions. If there is a specific additional experiment that would help convince you, we can attempt to provide the results within the remainder of the rebuttal time.

Reviewer_aS9QOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Contribution attribution: At comparable measured runtime, how do DAPS with landing-only refinement and DAPS with mean-curvature landing compare with DAPS+IIL on representative FFHQ tasks? A stochastic mean-curvature implementation would be a useful control. Since Eq. (7) expresses mean curvature as an expectation of directional curvature, comparing an independent trace probe with the same-noise coupling used by IIL would also help identify the source of the benefit.

Distributional validation: On the synthetic manifolds with known target distributions, can you report mode probabilities or a distributional discrepancy using multiple test functions, alongside the potential expectation? How sensitive are these results to initialization and simulation horizon? These checks would help distinguish improved geometric adherence from accurate exploration of the intended target.

Critical strengths

The formulation provides an interpretable mechanism for combining heterogeneous requirements. Expressing requirements through explicit tolerances and adapting their multipliers is an appealing alternative to manually balancing several guidance losses. The connection between the KL-constrained problem, exponential tilting, and reverse diffusion gives the algorithm a coherent mathematical motivation.

The method demonstrates a useful memory–quality trade-off. Avoiding denoiser backpropagation approximately halves peak allocated memory relative to several guidance baselines in the reported experiments. The semantic–photometric results also show strong joint accuracy, and the inverse-problem experiments demonstrate that the approach can accommodate additional constraints such as total variation. Using an independent classifier for semantic evaluation is a valuable experimental choice.

Critical weaknesses

he theoretical guarantee does not establish the central constraint-enforcement claim for the implemented sampler. Theorem 4.4 bounds the average dual-objective gap along the trajectory, with a residual bias This bias need not vanish even when the stated guidance and kernel errors vanish, because posterior/readout approximation terms remain. The theorem therefore does not establish convergence to the optimum, feasibility of the terminal output distribution, or a quantitative terminal constraint-violation bound. The practical configurations also depart from the analyzed algorithm through unbounded multiplier projection and, in some experiments, DDIM sampling and Adam updates.

These distinctions matter because the paper emphasizes enforcing requirements rather than merely encouraging them. Appendix E.5 documents meaningful failures: only 65% of nonlinear-deblurring outputs meet the observation tolerance, and adding TV to desolarization reduces observation feasibility to 0%. Although moment constraints do not require every sample to be feasible, the paper needs a clearer connection between its theoretical result and the constraint satisfaction claimed for practical outputs.

Questions for authors

Controlled comparison: How do the semantic–photometric results change when the baselines receive comparable loss normalization and tuned relative weights, or tolerance-aware penalties? An ablation comparing adaptive multipliers with fixed or scheduled weights using the same gradients and sampler would directly test the contribution of the dual updates.

Approximation and robustness: Section 4 states that omitting the denoiser Jacobian has little performance impact. Can you provide a direct comparison with and without this Jacobian under otherwise identical settings, together with sensitivity to the dual learning rate? This would clarify the accuracy–memory trade-off and the extent of task-specific tuning required.

6 comments

We thank the reviewer for their insightful comments. Before replying, we would like to point out that their Paper Summary may refer to a different submission, as it mentions "DAPS+IIL" and "mean-curvature methods," neither of which appear to apply to our work. Please let us know if this is not the case.

Next, we address each of the reviewer’s comments point by point. We are preparing a revised version of the manuscript, including new experiments and remarks (we thought it preferable to give priority to our responses). We will let them know as soon as the revised version is ready.

Weakness 1.1

We address the reviewer’s point in two steps: first, we discuss what Theorem 4.4 establishes and then the “bias.”

Scope of the theorem: The reviewer has a point that Thm 4.4 does not directly establish “convergence to the optimum, feasibility of the terminal output distribution, or a quantitative terminal constraint-violation bound.” Still, it provides a meaningful theoretical justification for the central optimization mechanism of PD-CDS and its guarantees are typical for dual ascent methods used in constrained optimization [see, e.g., (Nedić & Ozdaglar, 2009; Bertsekas, 2015)]. Indeed, recall from Prop. 3.1 that the constrained sampling problem (PI)~\tilde{\text{(PI)}} is equivalent to solving (DI)~\tilde{\text{(DI)}}, i.e., to maximizing the dual objective d~(λ)\tilde{d}(\lambda). In fact, as Remark 4.1 explains, the solution λ~⋆\tilde{\lambda}^\star of (DI)~\tilde{\text{(DI)}} fully characterizes the optimal (and feasible) distribution π~⋆\tilde{\pi}^\star, which can then be sampled using the corresponding Doob hh-transform in (7) (see Corollary 3.3). Hence, approaching λ~⋆\tilde{\lambda}^\star is not vacuous.

What is more, since −d-d is convex (line 195), we can use Jensen’s inequality to write Thm 4.4 directly in terms of the dual variables. To do so, note that if d~\tilde{d} is mm-strongly concave, then

\leq \tilde{d}(\tilde{\lambda}^{\star}) - \tilde{d} (\bar{\lambda}_{T}) \leq \tilde{d}(\tilde{\lambda}^{\star}) - \frac{1}{T} \sum_{t=1}^{T} \mathbb{E} \big[ \tilde{d} (\lambda_t) \big] \quad \text{for} \quad \bar{\lambda}_{T} = \frac{1}{T} \sum_{t=1}^{T} \mathbb{E} \big[ \lambda_t \big].$$ The last term is exactly what is bounded by Thm 4.4 (and more precisely, Thm D.1 in Appendix D). From Prop 3.1, the Hessian of $\tilde{d}$ is $-\text{Cov}(g)$ under $\tilde \mu_0^\lambda$, which is negative-definite under mild conditions (e.g., when there exists no nonzero linear combination of $g_i$ that is constant $\tilde \pi$_0-almost everywhere). Alternatively, by Fenchel duality, $\tilde{d}$ is strongly concave whenever $g$ and the KL divergence are smooth (i.e., have Lipschitz continuous gradients) [see, e.g., Kakade et al., “On the duality of strong convexity and strong smoothness” or (Bertsekas, 2009)], which in our case reduces to assuming that the distributions involved in (PI) are bounded away from zero. This corollary illustrates how Thm 4.4 can be used to show that PD-CDS actually approximates the Lagrange multiplier multiplier $\tilde{\lambda}^\star$ (at a rate of roughly $1/\sqrt{T}$ and with the same approximation-dependent bias as in Thm 4.4). If the reviewer considers that this corollary provides more clarity on the scope of Thm 4.4, we can find space to include it in the main part of the paper for the camera-ready version. Hence, we believe that the guarantees in Thm 4.4 (and more formally in Thm D.1) are not trivial and that they do in fact point to why PD-CDS actually works. That being said, as we explain in the conclusion, we expect additional, stronger results to be achievable by combining our analysis with those specific to DDPM, e.g., (Chen et al., 2023; Silveri & Ocello, 2025; Conforti et al., 2025). **Bias factor C**: While the bias term may be hard to eliminate, it is certainly possible to reduce it. This is indeed a point that we do not make sufficiently clear in the current version of the manuscript (and that we are addressing in our revision). The residual bias is summarized in Thm 4.4 as $C = O(R(1+\varepsilon_h+\varepsilon_{GD}))$ (its explicit form is relegated to Thm D.1 in Appendix D). It depends on three terms. Two come from Assumption 4.3, namely $\varepsilon_h$ and $\varepsilon_{GD}$. The former can be reduced using a more accurate approximation of the Doob $h$-transform as in (Zhu et al., 2026) or (Moufad et al., 2025). The latter, by refining the reverse diffusion discretization. The final term $O(R)$ is in fact a crude approximation used to simplify the exposition in the main text. Indeed, comparing (14) from Thm 4.4 to (18) from Thm D.1, we notice that it incorporates two approximations introduced in Section 4, namely, - **Posterior/Tweedie approximation (eqs. 9-10)**, i.e., the error of evaluating the constraint at the conditional mean instead of computing its conditional expectation; and - **Readout approximation (eqs. 10-11)**, i.e., the error of using an approximate guided score in Tweedie's formula. The readout error can also be reduced by improving the Doob $h$-transform approximation and the posterior error could be reduced by using better estimates of the conditional expectation of the constraint functions (e.g., with higher-order approximations).

Weakness 1.2

The fact that some experimental configurations depart from Algorithm 1 or the assumptions of Thm 4.4 showcase the versatility of PD-CDS in practice (that can also be instantiated with DDIM samplers or Adam) and its robustness to the assumptions used in our theoretical results. These include the upper bound on the projection of the dual variables. In fact, bounds derived from Slater’s condition are typically considerably larger than any value that λt\lambda_t could achieve in a few thousand steps. What is more, in the not-Young + contrast task in which Adam is used, replacing it by the standard dual ascent from Algorithm 1 yields similar results with 87% (94%) joint accuracy and a KID of 0.0372 (0.0548) for high (low) contrast. Still, the reviewer has a point that we could have been clearer about these points in our experimental description and we will do so.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Note to reviewers

We have updated our manuscript with the new experiments and remarks raised during the rebuttal. All modifications are highlighted in blue. In particular, we included more details on the bias in line 371 and the Jacobian in line 299 (experiments can be found in appendix E4). We have also highlighted the fact that our analysis is for Algorithm 1 and not the idealized scheme in (8), underscoring the approximations and errors we control (line 352). We have noted and justified why some of our experiments depart from the setting of Theorem 4.4 and Algorithm 1 (line 395) and replaced "NFE" by "number of diffusion steps," as this more clearly reflects that guided diffusion methods can incur substantially different computational cost per steps (Tables 4, 10, and 11 as well as explanations in the text). Most new experiments were included in Appendices E.4 .

~Alejandro_Ribeiro1OpenReview ↗

The citation of (Hadou et al. 2025) (https://arxiv.org/pdf/2606.17192) is disingenuous

The citation of (Hadou et al. 2025) (https://arxiv.org/pdf/2606.17192) is disingenuous. Algorithm 1 in this paper is the exact same algorithm proposed in (Hadou et al. 2025). Granted, in this paper the part of the score associated with the constraint (sλs_\lambda) is approximated by the weighted sum of constraint gradients whereas in (Hadou et al. 2025) it is learned. But this is the standard score approximation used in guidance. I.e., we know that sλs_\lambda requires integration over the posterior but it is common in practice to approximate it by the constraint gradient.

Furthermore, the theoretical analysis in this paper assumes that sλs_\lambda is the true score (Assumption 4.3). Thus, the analysis in this paper applies to the algorithm proposed in (Hadou et al. 2025), not the algorithm proposed here. In particular, Theorem 4.4 in this paper is equivalent to Theorem 1 in (Hadou et al. 2025). It is a weaker version in fact, because it applies to time averages of multipliers instead of the terminal multiplier λT\lambda_T.

This paper is also different in the use of KL divergences instead of entropies, but this is a trivial distinction. Neither the algorithm nor the analysis changes.

For what is worth, I do not oppose the publication of this paper. But if this paper is to be published, (Hadou et al. 2025) should be acknowledged in the introduction, not buried as an Appendix citation. Moreover, the introduction of the paper should clearly state that the algorithm is the same except for a modification in how the score sλs_\lambda is estimated and the application of the algorithm to enforce constraints on a pretrained diffusion model.

These may be good contributions as extensions of (Hadou et al 2025), but pretending that the algorithm and analysis are meaningfully different is wrong.

We appreciate the comment that makes a very good point: we should indeed be more precise about the distinctions with (Hadou et al., 2026) in the main text [the comment has a typo: the paper appeared in arxiv in 2026, first in June (v1) and then October (v2)]. This typo is consequential: our work is certainly not an extension. The works are, if anything, simultaneous contributions [by ICLR guidelines, papers are contemporaneous “if they are published within the last two months […] (i.e., at a peer-reviewed venue).”] In fact, they explicitly state that “authors are not required to compare to papers solely on arXiv.” Still, though we became aware of (Hadou et al., 2026, v1) late in our writing process, we discussed it in our related work section. But we agree that we could have written a more precise comparison. We do so below and will update our introduction with a more explicit discussion of (Hadou et al., 2026).

To be clear, PD-CDS is not the exact same algorithm proposed in (Hadou et al., 2026):

  • PD-CDS does not require training a new diffusion model. It is therefore training-free, a point we make extensively since we believe it is a crucial advantage of PD-CDS. This is in contrast to (Hadou et al., 2026) (see their Algorithm 1 in Appendix B.1 of v1), which in this sense is akin to classifier-free guidance (CFG);
  • What is more, the model in (Hadou et al., 2026) is trained not from samples, but from an explicit target distribution [namely, EE in (5) of (Hadou et al., 2026 v1)], as can be seen in line 19 of Algorithm 1 and (13) in Appendix B.1. Their problem is therefore closer to MCMC than generative learning;
  • PD-CDS does not use an expectation when updating the dual variables λ\lambda (step 5 of Algorithm 1 in our manuscript). In that sense, (PDI-D) from (Hadou et al., 2026) is comparable to the dual update in (8) from our manuscript. In practice, (Hadou et al., 2026) approximates this expectation using a batch of samples.

Contrary to what the comment seems to imply, these distinctions are not trivial. In traditional guidance, the “weighted sum of gradients” is based on conditional probabilities and Bayes rule (Doob’s hh-transform). We prove an extension that shows that a similar transform can solve the constrained sampling problem (PI) (Prop. 3.2 and Corollary 3.3).

Due to these differences, Theorem 4.4 is not equivalent to Theorem 1 in (Hadou et al., 2026). In particular, Assumption 4.3 assumes that the unconditional score sθs_\theta is exact, but not the “guidance” ∇xlog⁡htλ(x)\nabla_x \log h_t^\lambda(x): our result explicitly accounts for the approximation errors in this Doob-like transform. We also account for the stochastic approximation in step 5 of PD-CDS [Theorem 1 in (Hadou et al., 2026) considers the update in expectation as per (PDI-D)]. Hence, Theorem 1 in (Hadou et al., 2026) does not apply to PD-CDS. What is more, (Hadou et al., 2026, v1) assumes that the accumulated errors are bounded in Lemma 4 [specifically, (82) and (83)]. These appear to have become Assumption 6 in v2 (October 2026). Our analyses provide dual convergence guarantees without these assumptions.

Finally, note that the results in (Hadou et al., 2026, v1) are also for “time averages of multipliers λˉ\bar{\lambda}” and it is only the v2 from October (after the ICLR deadline) that gives results for the terminal multiplier λT\lambda_T. We feel this is also a consequential point. These results hold for a different algorithm (that retrains diffusion models) and under stronger assumptions than the ones from our Theorem, namely Assumptions 3-6. In fact, this includes the assumption on the strong concavity of the dual function that we mention in our response to Reviewer aS9Q.

We hope our response has clarified what we consider to be substantial differences between the algorithms and their analyses. On our side, we will include this detailed comparison in our related works and also ensure that it is given a more appropriate place in our introduction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.