acceptodds
Under review as a conference paper at ICLR 2027

Constrained Diffusion from Binary Evaluations via Learned Doob h-Functions

Abstract

Conditional generation under constraints that can only be checked on a finished sample, such as a molecular property filter, remains challenging for diffusion models: most guidance methods require differentiable measurement operators, paired conditional data, or rollouts of the base model at every step. We present a framework grounded in the Doob -transform that reduces conditional generation to learning a single function, the probability that a noisy sample denoises into the constraint set. We learn this function from binary clean-sample evaluations alone, by squared-error regression on forward-noised labels; the base model is frozen, one lightweight head is trained per constraint, and noisy-state classifier guidance arises as a special case. The exact guidance step fades where the constraint signal provably vanishes and carries a factor of the inverse probability across samples, which normalized guidance discards: with the schedule and mean step size held fixed, equalizing the step costs 17.0 to 18.5 points of constraint satisfaction on CIFAR-10 at comparable FID, across independently trained estimators. Guidance is more than cheaper per satisfying sample than rejection sampling at a 1.7% base rate, raises a Lipinski-type property filter on QM9 from 54.3% to 96.8%, and lifts satisfaction on Stable Diffusion 1.5 from 7.2% to 60.0% at one sixteenth of the cost of derivative-free selection, which reaches 9.0%. These results demonstrate that binary evaluations of clean samples are sufficient supervision for constrained generation across pixel, latent, and molecular diffusion.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
diffusion models, guidance, Doob h-transform, constrained generation, backward Kolmogorov equation, posterior sampling, conditional generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zHSDOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is about diffusion generation under constraints available only as binary evaluations of final samples. It learns a Doob h-function on noised labeled data and guides a frozen diffusion model with its log-gradient. Experiments on MNIST, CIFAR-10, QM9, and Stable Diffusion 1.5 evaluate constraint satisfaction, sample quality, and sampling efficiency.

Critical strengths
  • Coherent formulation and analysis. The paper connects binary supervision, the Doob h-function, and the resulting guidance rule together. Method design and experiments are coherent, providing evidence to support the claims.
Critical weaknesses
  • The experiment conditions are too simple to establish effectiveness on more complex constraints. The experiments primarily use digit or class labels, low-level image statistics, grayscale conversion, and basic molecular property thresholds or ring presence. The composition experiments also offer limited evidence about strongly dependent constraints.
  • The incremental contribution over existing classifier guidance is limited. The paper's idea is already encompassed by classifier guidance. Standard classifier guidance also retains the guidance magnitude. The paper doesn't establish a sufficiently substantial advance over the existing formulation to support acceptance.
Questions for authors
  1. Clarification of the theoretical contribution: Which part of the method establishes a new result beyond the existing classifier-guidance and Doob h-transform formulations? What additional conclusion does it support about the proposed sampler?

  2. Generality of the central empirical finding: Table 2 uses standard classifier guidance to demonstrate the benefit of retaining magnitude differences across samples, but only for two simple CIFAR-10 predicates. Since this analysis is a central contribution, can you demonstrate the same benefit for a more complex constraint, with the schedule and mean step size matched and sample quality comparable?

5 comments
AuthorsOpenReview ↗

Response to Reviewer zHSD (1/2): contribution and class constraints

We thank the reviewer for the careful reading. We address the contribution of the paper first, then the experiments on more complex constraints, including a dependent conjunction and a molecular constraint, and the composition of dependent constraints.

1. Contribution beyond classifier guidance and the Doob h-transform

The distinction that matters is between the established guidance formula and the question this paper studies. Standard classifier guidance already uses the unnormalized logarithmic gradient ∇log h. The paper isolates what the relative magnitude of that gradient contributes to hard-constraint satisfaction, independently of its direction, of the timestep schedule and of the average guidance budget.

The finding is quantitative and bears directly on sampler design. In the matched-budget experiment of the paper, giving every sample an equal step while keeping the mean step at each timestep lowers satisfaction by 17.0 to 18.5 points at comparable FID_c, and restoring only the inverse-probability weighting recovers a large part of the loss. During the discussion period we extended the comparison to 32 configurations and three estimator families, including the independently trained estimator of Spectral Guidance. The allocation of magnitude is therefore not an incidental feature of our architecture. It is a measurable aspect of conditional guidance, and its effect can be tested separately from the quality of the estimator and from the overall guidance strength.

The theory characterizes the hard-constraint setting in which this takes place. Theorem 1 gives the bounded-solution characterization of h for a discontinuous terminal indicator and identifies the exact conditional dynamics: with the exact score, the exact h and the conditional initial law, the correction at γ = 1 is the Doob-transformed process. Proposition 2 quantifies how the constraint signal contracts under forward noising, and Proposition 3 distinguishes the data-law target from the model-law target. These results establish the conditional correction and its scope, and the controlled experiments establish what happens when its magnitude is reallocated with learned estimates.

For the sampler, the practical conclusion is that rules which normalize, clamp or otherwise modify the guidance magnitude should be assessed for how they distribute the correction across samples, and not only for their average strength. The matched-budget protocol gives a concrete way to do this, including for estimators developed independently of ours.

2. Matched-budget experiment on class constraints

We ran the same matched-budget test on two semantic constraints, the CIFAR-10 classes cat and ship, judged by a separately trained classifier. The protocol is unchanged: the three rules use the same classifier gradient direction, and the equal and 1/ĥ rules give each sample a step whose mean is the standard rule's mean step at every t (1,000 samples per run; satisfaction % / FID_c).

ClassγSampling seedStandardEqual1/ĥ
Cat2042.8 / 62.241.6 / 61.242.0 / 62.7
Cat5064.3 / 58.660.3 / 58.163.0 / 58.8
Cat5168.0 / 60.065.9 / 58.767.1 / 59.9
Cat10076.4 / 63.272.0 / 61.675.5 / 62.8
Ship2054.9 / 51.251.0 / 55.651.3 / 52.5
Ship5080.7 / 40.272.2 / 41.177.2 / 40.2
Ship5179.6 / 41.071.4 / 41.975.0 / 41.1
Ship10088.6 / 41.581.4 / 41.183.7 / 41.4

The matched-budget effect extends to semantic class constraints. The equal step lowers satisfaction in all eight runs, with FID_c within 1.6 of the standard rule in seven of them. The effect is strongest on ship: at γ = 5 the loss is 8.5 and 8.2 points in the two sampling seeds, while FID_c changes by less than one point, and at γ = 10 it is 7.2 points. On cat it is 1.2 to 4.4 points, in the same direction in every run. The inverse-probability weighting recovers part of the loss in every run.

These results take the original observation beyond low-level image statistics. Together with two further pixel predicates, further guidance scales, coarser samplers, our own head, the Spectral Guidance estimator, the dependent conjunction and the molecular filter reported below, the comparison now spans 32 configurations, 30 on CIFAR-10 and 2 on QM9, against 6 in the submission, with lower satisfaction under the equal step in every one. The effect is thus not specific to the original predicates or to our head, and its size varies with the constraint.

AuthorsOpenReview ↗

Response to Reviewer zHSD (2/2): dependent conjunction and molecular constraint

3. Composition of strongly dependent constraints

This experiment addresses a separate aspect of generality: whether independently trained heads remain useful when the two constraints are strongly dependent. Red-pixel and low saturation co-occur in 1.28% of CIFAR-10 training images, against 6.69% under independence. Guiding with the sum of the two separately trained log-heads, with no joint training:

γJoint satisfaction (%)Red-pixelLow saturationProduct of guided marginals
00.531.415.14.7
230.484.239.533.3
562.5 ± 2.097.064.062.1
1083.4 ± 0.397.683.981.9

(3 seeds × 1,000 samples at γ = 5 and 10; seed 0 otherwise.) Joint satisfaction rises from 0.5% without guidance to 83.4% at γ = 10, and the result is reproduced over three seeds at the two larger scales. Composition is therefore effective for this strongly negatively dependent pair, with heads trained once per constraint and combined at sampling time. Exact additive composition requires conditional independence given the noisy state, which is not assumed here; the experiment shows successful joint satisfaction for one dependent pair and is not a guarantee for arbitrary constraints.

4. Matched-budget experiment on the dependent conjunction

To test the allocation effect on a compound constraint, we ran the matched-budget comparison on this same conjunction. The direction is the gradient of the summed log-heads, the mean step at each timestep is measured along the standard-guidance trajectories, and the 1/ĥ rule uses the product of the two heads (3 sampling seeds × 1,000 samples; joint satisfaction %, mean ± standard deviation, and mean FID_c against the test images that satisfy both predicates).

γStandardEqual step1/ĥ-weighted
562.5 ± 2.0 / 118.657.3 ± 2.2 / 118.861.6 ± 2.0 / 118.5
1083.4 ± 0.3 / 131.176.1 ± 0.4 / 131.580.4 ± 1.1 / 133.6

The equal step lowers joint satisfaction in all six seed-level comparisons, by 5.2 points on average at γ = 5 and by 7.3 points at γ = 10, where all three seeds show losses between 7.0 and 7.6 points, while FID_c stays within 0.5 of the standard rule. Restoring only the inverse-probability weighting recovers 82% of the mean loss at γ = 5 and 59% at γ = 10. Because the two events are dependent, the product of the heads is not assumed to equal the true joint conditional probability; the recovery is an empirical result of the controlled reweighting. The allocation effect therefore holds for a rare, negatively dependent conjunction guided by two separately trained heads, with the same controls as in the original experiment.

5. Matched-budget experiment on a molecular constraint

To take the comparison beyond image predicates, we ran it on the Lipinski-type molecular filter used in the paper. This is a programmatic molecular property constraint: an RDKit program on decoded molecules, with no gradient and no noisy-state evaluation. The sampler is the one used in the paper, and the three rules are the same as before, with the mean step at each timestep measured along the standard-guidance trajectories (3 sampling seeds × 1,000 molecules; satisfaction %, mean ± standard deviation, with mean uniqueness / novelty % in parentheses; validity is 100% in every run).

γStandardEqual step1/ĥ-weighted
287.4 ± 0.5 (93.8 / 95.4)86.3 ± 1.2 (93.1 / 96.3)87.8 ± 0.7 (93.2 / 96.1)
593.7 ± 0.4 (89.5 / 94.4)91.0 ± 0.7 (89.5 / 95.3)93.5 ± 0.9 (89.2 / 94.7)

The equal step is lower in all six seed-level comparisons. At γ = 5 the loss is 2.7 points on average, between 1.5 and 3.3 points across the three seeds, and the inverse-probability weighting recovers over 90% of it. At γ = 2 the loss is 1.1 points, in the same direction in all three seeds. Validity is 100% under all three rules, and uniqueness and novelty differ by less than one point between them, so the loss in satisfaction is not offset by a gain in validity, uniqueness or novelty. The allocation effect therefore carries over to a programmatic molecular constraint in a different generative domain.

AuthorsOpenReview ↗

Follow-up to Reviewer zHSD: complex constraints and an independent estimator

We thank the reviewer again for the thoughtful questions, which helped shape the additional experiments. Two findings bear directly on the concerns raised.

On generality beyond simple constraints: we extended the matched-budget evaluation to more challenging settings, preserving the reference timestep-wise mean-step schedule, at comparable sample quality. On the semantic class ship, equal-step guidance lowers satisfaction by 8.5 and 8.2 points across two seeds at γ = 5. On a strongly dependent conjunction, it lowers joint satisfaction from 83.4 ± 0.3% to 76.1 ± 0.4% at γ = 10, with FID_c differing by less than 0.5. On the QM9 molecular property filter, satisfaction decreases from 93.7 ± 0.4% to 91.0 ± 0.7% at γ = 5, with validity remaining at 100%.

On the contribution beyond existing classifier guidance: we also repeated the matched-budget experiment using the independently developed Spectral Guidance estimator, trained with a different objective. Equal-step guidance again lowers satisfaction in all four tested configurations (1.8–5.9 points, across three sampling seeds). Together with the more complex constraint experiments, this demonstrates that the magnitude-allocation effect extends beyond our particular estimator and the original image predicates. Our contribution is to isolate this effect under a controlled guidance budget and provide a protocol through which other conditional-guidance methods can evaluate it.

We sincerely appreciate the reviewer's feedback, which helped strengthen the scope and experimental validation of our findings. We will incorporate these results and their implications into the revised manuscript.

Reviewer_5LBYOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies diffusion guidance toward a binary label, computed by a non-differentiable checker on clean samples. It frames guidance as a Doob hh-transform with ht(x)=p(y=1∣xt)h_t(x) = p(y=1 \mid x_t) and learns hh by regressing clean labels on forward-noised inputs. The authors characterize hh and show that its variance contracts with noise. They argue that the guidance step should not be normalized since the unnormalized ∇h/h\nabla h / h shrinks at high noise and pushes low-probability samples harder. Experimental setup covers CIFAR-10 pixel predicates, MNIST, QM9 and Stable Diffusion 1.5. The main experiment compares the learned head against a matched noisy classifier under standard and normalized guidance, and includes an ablation that fixes the average guidance strength per timestep and varies only how it is distributed across samples. Further experiments compare against rejection sampling and SVDD in cost per satisfying sample, and test composition of independently trained heads on MNIST and QM9.

Critical strengths
  • The paper is honest in delimiting its contribution, stating that the proposed estimator coincides with classifier guidance (and Table 1 confirms equivalent performance).

  • The matched-budget experiment in Section 6.2 is a well-controlled ablation that holds the schedule and mean step fixed and varies only per-sample allocation, with results replicated across three classifiers.

  • Implementation details are thorough.

Critical weaknesses
  • Presentation: I found the paper hard to follow. The first paragraph jumps straight to the score decomposition before the problem is even posed. The reader is never given any context and there's no mention that the method is, in fact, standard conditional guidance with a binary label.

  • Novelty: Since the estimator is the same as classifier guidance, the contribution is essentially the Doob-h analysis and the claim that the guidance step should not be normalized. I think most of the theory is standard. However, the paper does not cite or discuss Spectral Guidance (Moreira et al., ICML 2026), which expresses guidance through a conditional expectation operator and also learns it from independently forward-noised copies of clean samples. It handles binary labels (which is the core application of this paper), and its spectral decay analysis yields the signal-contraction behavior of Prop. 2. The authors should position their analysis relative to this work and compare against it empirically, including how its guidance rule handles step magnitude.

  • Claims. In Section 6.1, at equal γ\gamma the normalized step is much larger than the unnormalized one, especially at high noise, so the FID gap reflects an unmatched push off the data manifold. Section 6.2 controls for this, but only at a single γ\gamma on two predicates. Moreover, on Stable Diffusion 1.5, the normalized rule at its best setting (65.1% at γ\gamma = 10, the smallest scale tested) outperforms the proposed method’s best at similar FID: the paper’s only large-scale test seems to contradict its core claim.

Questions for authors
  • How does the method compare to Spectral Guidance on the same binary labels? Plus, how does Prop. 2 relate to the spectral decay / optimal guidance window analysis from that paper?
  • What does the normalized rule give on SD 1.5 for γ\gamma < 10?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 5LBY (1/4): novelty, and comparison with Spectral Guidance

We thank the reviewer for the detailed review. We address novelty first, then the experimental points in the order raised, and then presentation.

1. Novelty and significance

The central contribution of the paper is a controlled finding about how conditional guidance is applied. The Doob h-transform and classifier guidance establish the correction ∇log h, and for binary labels their target coincides with ours. Standard classifier guidance uses this correction with its magnitude, and other rules rescale it; in both cases the allocation of magnitude across samples is confounded with the overall guidance strength. What the paper isolates is the consequence of keeping or discarding the relative magnitude of that correction across samples, with the strength held fixed. Two samplers can use the same gradient directions, the same timestep schedule and the same average step size, and still reach substantially different constraint-satisfaction rates.

The comparison is direct. Starting from one estimated guidance field, Section 6.2 measures the mean step at each timestep along the standard-guidance trajectories, and then gives every sample an equal step of that size, leaving its direction unchanged. Satisfaction falls by 17.0 to 18.5 points at comparable FID_c (Table 2), and restoring only the relative 1/ĥ weighting, at the same reference mean-step schedule, recovers a large part of the loss. The extensions now cover 32 configurations and three estimator families, including the independently trained estimator of Spectral Guidance (points 3 and 6). The effect is therefore not explained by applying more guidance, by a favourable choice of γ, or by our particular head.

This gives later work a testable design variable and a protocol for measuring it: when a method normalizes or rescales conditional guidance, the distribution of step magnitudes across samples can be evaluated separately from the overall guidance budget. Prior work constructs, learns, schedules or rescales the conditional correction (Guo et al., 2026; Chang et al., 2026; Yang et al., 2024; Ma et al., 2024); our experiments quantify what changing its allocation across samples costs.

The theory characterizes the hard-constraint target on which this study rests, and it is specific to that case. Theorem 1 treats a discontinuous terminal indicator, which the existing guarantees for forward-noised regression exclude because they require a weight bounded away from zero (Chang et al., 2026). Proposition 2 gives monotone contraction of the constraint signal with an explicit bound under a finite second moment, where the spectral analysis of Spectral Guidance gives a limit under compact support (point 4). Proposition 3 distinguishes the data-law target from the model-law target of rollout methods. The theorems identify and analyse the conditional correction, and the matched-budget experiments establish the effect of reallocating it.

The formulation is also practical. Binary evaluations of finished samples guide a frozen generator without differentiating the verifier or retraining the model: a molecular property filter rises from 54.3% to 96.8%, and for a rare image predicate guidance is 53 to 55 times cheaper per satisfying sample than rejection.

2. Empirical comparison with Spectral Guidance

We compared the two methods on the binary labels of Sections 6.1 and 6.2, using the authors' released code and their sampler. Each method's guidance strength was selected on sampling seed 0, and the comparison is reported on seeds 1 and 2, which play no part in the selection (mean over the two evaluation seeds × 1,000 samples):

PredicateMethodStrengthSatisfaction (%)FID_cFID_fs/sample
Red-pixelSpectral Guidance1.599.8576.0799.180.039
Red-pixelOurs0.2599.7051.3263.310.025
Upper luminanceSpectral Guidance1.099.5049.4185.700.039
Upper luminanceOurs0.2599.9543.9666.210.025

At satisfaction within 0.5 points, our FID_c is 24.8 lower on red-pixel and 5.5 lower on upper luminance, and sampling takes 35% less time. Each evaluation seed is within 0.1 of the mean in satisfaction and within 1.3 in FID, and including the tuning seed changes each mean by at most 0.3.

AuthorsOpenReview ↗

Response to Reviewer 5LBY (2/4): protocol of the Spectral Guidance comparison, and step magnitude

2 (continued). Protocol of the comparison

Since no trained network is released, we trained their network with the authors' code and CIFAR-10 configuration on the same base model as ours, and confirmed that it shows the spectral transition they report near t ≈ 400. Both methods then run in their sampler with their settings (DDIM, 100 steps): theirs with the released step rule, ours with ∇log h_θ from the heads of the paper. Our copy of their sampling loop reproduces theirs exactly.

Each method was tuned in the same way. We swept the guidance strength on sampling seed 0 (8 values for ours, 9 and 10 for Spectral Guidance), and selected the strength with the lowest FID_c among those reaching at least 99.9% satisfaction on that seed. Each method has one trained network, so the variation across seeds reflects sampling variability. FID_c is computed against the CIFAR-10 test images that satisfy the predicate (2,704 for red-pixel, 2,416 for upper luminance) and FID_f against the 10,000 test images, as in Appendix B.2 and B.4.

Over the whole sweep on seed 0, ours has the lower FID_c on red-pixel at every satisfaction level their method reaches, and on upper luminance ours attains the lower FID_c at every satisfaction level above 99.5%, the regime in which a constraint is meant to be enforced.

The time is total sampling time per sample, measured for both methods in the same loop on the same GPU; it includes every base-model evaluation and the guidance gradient. Guidance itself adds 0.002 s per sample with our head against 0.016 s with their network, since our head is more than ten times smaller (0.68M against 9M parameters).

We also evaluated the constraint-prediction quality of the two estimators. On 10,000 forward-noised test images, our head achieves higher AUROC at all six evaluated noise levels (t = 10, 100, 300, 500, 700 and 900) on both predicates (AUROC at t = 100: 0.987 against 0.895 on red-pixel, 0.999 against 0.984 on upper luminance). Their network learns one label-agnostic basis and projects each label onto its leading modes, whereas our head is regressed on the constraint itself. This comparison is between the two complete methods; point 3 isolates the step rule.

3. Step magnitude in the Spectral Guidance sampler

Their released sampler computes ∇p̂/p̂ and rescales each sample's gradient to a norm in [0.1, 0.3] (clamp bounds 0.1 and 0.3 in the released code). Across samples the step therefore varies by at most a factor of 3. We ran the test of Table 2 on their trained estimator, comparing the unrescaled log-gradient, an equal step and an inverse-probability-weighted step at the same schedule and mean step (3 seeds × 1,000 samples; satisfaction %, mean ± standard deviation, and mean FID_c). This controlled test is separate from their released clamped sampler, which is the one evaluated in point 2; its purpose is to check whether the allocation effect persists when the estimator is changed.

Predicate, strength∇p̂/p̂Equal step1/p̂-weighted
Red-pixel, 0.2578.97 ± 1.13 / 59.073.10 ± 1.34 / 56.977.33 ± 1.52 / 58.9
Red-pixel, 0.594.53 ± 0.57 / 66.589.27 ± 0.78 / 63.692.70 ± 0.91 / 65.7
Upper luminance, 0.2595.60 ± 0.22 / 42.292.53 ± 2.97 / 43.296.83 ± 1.75 / 43.3
Upper luminance, 0.599.87 ± 0.05 / 46.098.07 ± 0.65 / 45.199.87 ± 0.12 / 46.0

The equal step has the lower mean satisfaction in all four rows (5.9, 5.3, 3.1 and 1.8 points). The difference exceeds the seed standard deviations in three of the four rows. FID_c of the equal step is within 3 of the unrescaled step in every row. The mechanism of Section 6.2 therefore transfers to an estimator of h that was developed independently and trained with a different objective. It is a property of how guidance is allocated, not of the estimator we trained.

AuthorsOpenReview ↗

Response to Reviewer 5LBY (3/4): Proposition 2 and the spectral analysis; guidance scales

4. Relation of Proposition 2 to the spectral analysis

The two analyses are connected by an identity. With TtT_t their conditional-expectation operator and (σt,k,ψt,k)(\sigma_{t,k}, \psi_{t,k}) its singular system, ht−hˉ=Tt(1C−hˉ)h_t - \bar h = T_t(1_C - \bar h), so

Var[ht(Xt)]=∑k≥2σt,k2⟨1C,ψt,k⟩2.\mathrm{Var}[h_t(X_t)] = \sum_{k \ge 2} \sigma_{t,k}^2 \langle 1_C, \psi_{t,k} \rangle^2 .

Our σ2(t)\sigma^2(t) is their spectrum weighted by the coefficients of the constraint, so the two analyses describe the same contraction from two sides: theirs for a reusable basis, ours for one constraint. Proposition 2 adds three guarantees. It gives monotonicity at every tt, where their Proposition A.7 gives the limit as αˉt→0\bar\alpha_t \to 0. It gives an explicit bound through the mutual information. And it requires only a finite second moment, where theirs assumes compact support. Being specific to the constraint, it bounds the signal that a given constraint retains at a given noise level. Separately, in our experiments the unnormalized drift concentrates its step at intermediate noise without a guidance window being set: its mean step peaks near t = 500 on red-pixel and t = 700 on upper luminance and is 4.4 and 3.7 times smaller at t = 999.

Our analysis also addresses questions that their spectral framework does not develop: the characterization of hh for a discontinuous indicator (Theorem 1), the distinction between data-law and model-law targets (Proposition 3), and, most centrally, the effect of reallocating guidance across samples, which point 3 tests on their independently trained estimator.

5. Comparison across guidance scales (Section 6.1)

To remove any dependence on the scale at which the two rules are compared, we extended the normalized rule down to γ = 0.1 and compared the two along their whole satisfaction–quality curves (seed 0, satisfaction % / FID_c):

  • Red-pixel. Normalized: 39.3 / 52.6, 50.2 / 51.0, 74.4 / 53.3, 95.1 / 61.7, 100 / 108.1. Ours: 55.4 / 49.7, 80.3 / 49.4, 97.7 / 50.6, 100 / 58.2.
  • Upper luminance. Normalized: 29.5 / 66.0, 44.7 / 56.3, 69.8 / 45.1, 96.8 / 43.2, 99.9 / 62.3, 100 / 80.3. Ours: 55.8 / 54.8, 90.9 / 45.0, 99.2 / 39.5, 100 / 43.4.

Our drift gives the better trade-off at every tested operating point on both predicates: for every point of the normalized rule there is a point of ours with satisfaction at least as high and lower FID_c. The gap is largest where it matters most, at high satisfaction. At 100% satisfaction our FID_c is 58.2 against 108.1 on red-pixel and 43.4 against 80.3 on upper luminance, and the normalized rule's first point above 95% on red-pixel is at FID_c 61.7, while ours is at 50.6 at 97.7%. The advantage reported in Section 6.1 therefore does not depend on the scale chosen for the comparison.

Reviewer_QEw1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents a guidance method using Doob h-transform. It shows that the magnitude contains information and also enable cheap guidance compared to other methods. However, this paper does not differentiate it with other guidance method like SAG/PAG or reinforcement learning methods.

Critical strengths
  1. The h-transform is theoretically sound and useful
  2. the inference cost is small
Critical weaknesses
  1. The paper is quite full of theory without more applications or comprehensive experiments on different datasets/baselines.
  2. need more discussions with other guidance methods like a. SAG, b. PAG, c. CFG++ and so on.
  3. Has author investigated any reinfocement learning methods like Flow-GRPO, DiffusionNFT etc. Those algorithms has a reward such as "make it red" and train the rollout of model (can be a fixed inference method such as a CFG schedule, a sigma shift, step sizes), to inclining more to the reward/condition. I think it has something similar to your work. Specifically the figure in your paper looks similar to another Google paper, doing reward optimization in the noise space.
Questions for authors
  1. need more discussions with other guidance methods like a. SAG, b. PAG, c. CFG++ and so on.
  2. Has authors investigated how this method can be used in distillation like DMD or PDD. In distillation the guidance method matters a lot. I am wondering if this technique will be suitable for better control over distillation
  3. Has author investigated any reinfocement learning methods like Flow-GRPO, DiffusionNFT etc. Those algorithms has a reward such as "make it red" and train the rollout of model (can be a fixed inference method such as a CFG schedule, a sigma shift, step sizes), to inclining more to the reward/condition. I think it has something similar to your work. Specifically the figure in your paper looks similar to another Google paper, doing reward optimization in the noise space.
5 comments
AuthorsOpenReview ↗

Response to Reviewer QEw1

We thank the reviewer for recognising the soundness and low inference cost of the approach, and for the pointers to related methods. We will add the corresponding discussion to the related-work section.

1. Setting and contribution

The paper studies generation under a hard constraint that is accessible only through a binary evaluation of a finished sample. The generator stays frozen, and one lightweight head per constraint is trained on labelled, forward-noised data, so its training targets require no base-model rollouts.

The correction ∇log h is shared with classifier guidance. The contribution is to characterize it for hard constraints and to isolate the effect of how its magnitude is allocated across samples. With the schedule and the mean step held fixed, giving every sample an equal step lowers satisfaction by 17.0 to 18.5 points at comparable FID_c (Section 6.2). We extended this to 32 configurations and three estimator families, including the independently developed estimator of Spectral Guidance.

2. Attention-based guidance and CFG++

SAG (Hong et al., ICCV 2023) and PAG (Ahn et al., ECCV 2024) use the denoiser's attention maps to improve sample quality, and CFG++ (Chung et al., ICLR 2025) improves guidance for conditioning the model already has. Our setting adds a requirement: enforcing a new property, specified by a binary check, that is absent from the model's conditioning space. SAG and PAG do not estimate the event probability for such a property, and our head supplies that correction. The methods are complementary: in our Stable Diffusion experiments the head is applied alongside text-based classifier-free guidance.

3. Reinforcement-learning fine-tuning

Flow-GRPO (Liu et al., NeurIPS 2025) updates a flow-based generator by online policy-gradient reinforcement learning. DiffusionNFT (Zheng et al., ICLR 2026) optimizes the generator on the forward process, contrasting positive and negative samples that the model itself generates.

Our approach also uses forward noising, but differs in the training data, in what is trained, and in deployment. We learn the event probability h_t from labelled clean data and the known forward kernel, with no reward-labelled samples from the generator, and the generator is unchanged. Each constraint adds a head of 675,969 parameters on CIFAR-10, and heads can be combined. The two routes are reward-driven optimization of the generator and a constraint-specific correction applied to a frozen one. Proposition 3 further distinguishes the data-law probability we learn from the model-law probability of the generator.

4. Additional reference

We could not identify the Google paper from the description. Could the reviewer kindly provide the reference? We will cite and discuss it.

5. Distillation

Distillation answers a different question from the one this paper studies. It trains a new student generator to reproduce a teacher sampler in fewer steps. Our setting keeps the generator frozen and trains no generator at all: the contribution is the conditional correction and how its magnitude is allocated. Distillation is therefore downstream of our result, not an alternative to it. The connection is direct: with an exact score and an exact h, Theorem 1 gives the constrained score as s(x, t) + ∇log h_t(x), which is the natural constraint-conditioned teacher score for DMD and related methods. An experiment on a distilled student would evaluate the distillation procedure, not the claims of this paper, which is why it is outside its scope.

6. Applications, datasets and baselines

The submission evaluates MNIST, CIFAR-10, QM9 with RDKit property filters, and Stable Diffusion 1.5, against classifier guidance under two step rules, rejection sampling, SVDD in two variants, and derivative-free selection. On QM9 a molecular property filter rises from 54.3% to 96.8%; on Stable Diffusion 1.5 satisfaction rises from 7.2% to 60.0%; for a rare image constraint, guidance is 53 to 55 times cheaper per satisfying sample than rejection.

During the discussion period we added two things. First, a comparison with Spectral Guidance (ICML 2026) using the authors' code: at high satisfaction (99.5 to 99.95%), our FID_c is 51.3 against 76.1 on red-pixel and 44.0 against 49.4 on upper luminance (evaluation seeds separate from the tuning seed), with 35% less sampling time. Second, the matched-budget experiment of Section 6.2 was extended from 6 to 32 configurations. On the QM9 Lipinski-type filter, an equal step lowers satisfaction by 2.7 points at γ = 5 over three seeds, with validity at 100% and uniqueness and novelty within one point. On a conjunction of two strongly dependent image predicates, it lowers joint satisfaction from 83.4 ± 0.3% to 76.1 ± 0.4% at γ = 10 at comparable FID_c. The general comment lists the configurations.

I don’t understand what prevents you from training a Lora with a binary reward with reinforcement learning. The RL can be any algorithm: flow grpo, diffusionft, refl etc reward: like “make the color red”. This can definitely satisfy consistency and you can compare with your methods for consistency. Regarding to initial noise optimization techniques, please refer to this NeurIPS 2025 paper, which is the Google paper I am referring to: https://arxiv.org/pdf/2508.09968

AuthorsOpenReview ↗

Reply to Reviewer QEw1: reinforcement-learning fine-tuning and Noise Hypernetworks

We thank the reviewer for the thoughtful follow-up and for identifying the relevant reference.

1. Binary-reward reinforcement learning and LoRA fine-tuning

We agree that training a LoRA with binary rewards through Flow-GRPO, DiffusionNFT, or related methods is a promising approach to constrained generation. Such methods can effectively optimize a generator toward desired properties.

Our work addresses a complementary and practically important question: how can we achieve effective constraint satisfaction using an existing pretrained generator, without changing its parameters, and what makes such guidance successful?

This is an important setting in its own right. Modifying a pretrained generator may be undesirable or impractical when the same model must serve many constraints, when updating its weights is not possible, or when preserving the original generator is important. Rather than adapting the generator itself, we investigate how much can be achieved through a lightweight, modular conditional correction applied at sampling time.

Our results demonstrate that this approach can be highly effective. We train constraint-specific heads of approximately 0.68M parameters on CIFAR-10 using binary-labelled clean samples and the forward-noising process, without reward-labelled generator rollouts or updates to the generator. Guidance strength can be adjusted at inference time without retraining, and independently trained heads can be composed to enforce multiple constraints. On a strongly dependent conjunction, we increase joint satisfaction from 0.5% to 83.4% without joint training. On Stable Diffusion 1.5, we improve satisfaction from 7.2% to 60.0% while keeping the pretrained generator frozen.

Beyond demonstrating effective constrained generation, our central scientific contribution is understanding how conditional guidance should be applied to achieve strong constraint satisfaction. Our controlled experiments establish that the allocation of guidance magnitude across samples is consequential even when the reference timestep-wise mean-step schedule is matched. Equalizing per-sample guidance steps reduces satisfaction by 17.0–18.5 percentage points in our original experiments, at comparable conditional FID.

We have now validated this finding across 32 experimental configurations and three estimator families, including the independently developed Spectral Guidance estimator. The effect persists across semantic constraints, strongly dependent compositions (7.3-point satisfaction difference at nearly identical conditional FID), and molecular generation on QM9 (2.7-point difference). Our direct comparison with Spectral Guidance further demonstrates practical effectiveness: at similarly high satisfaction, our method achieves lower conditional FID and approximately 35% lower sampling time in the evaluated settings.

We therefore view RL-LoRA as a valuable approach to adapting a generator, while our work develops and systematically studies a complementary capability: controlling a fixed pretrained generator through principled, effective, and composable conditional guidance. Our baseline comparisons focus on this setting to isolate the contribution of the guidance mechanism and its magnitude allocation. We agree that a broader comparison with reward-based generator adaptation would also be informative.

2. Connection to Noise Hypernetworks

We thank the reviewer for identifying Noise Hypernetworks (Eyring et al., NeurIPS 2025). We agree that this is a closely related and valuable contribution, especially because both approaches aim to steer generation while preserving the pretrained generator.

The methods offer complementary mechanisms for achieving this goal. Noise Hypernetworks learns an auxiliary network that modifies the initial noise distribution of a frozen, distilled generator using differentiable reward objectives. Our method instead learns a time-dependent conditional event probability from binary-labelled clean samples and applies its logarithmic gradient throughout denoising. This allows us to address hard, non-differentiable constraints without requiring reward gradients through the generator, as demonstrated by our programmatic RDKit molecular-property constraints.

Our approach also has a distinct theoretical foundation. The Doob h-function formulation characterizes the exact conditional dynamics for a hard terminal event under idealized assumptions, providing a principled target for the learned guidance.

We appreciate the reviewer bringing both of these connections to our attention. We will update the Related Work section to discuss binary-reward RL fine-tuning, including Flow-GRPO and DiffusionNFT, and incorporate Noise Hypernetworks, clarifying the complementary objectives, mechanisms, and settings of these approaches relative to ours.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response (1/2): contribution and matched-budget study

We thank the reviewers for their time and for a discussion that prompted several new experiments. In response to the reviews we added three sets of evidence: a direct comparison with Spectral Guidance using its released implementation; an extension of the matched-budget study from 6 to 32 configurations and three estimator families, now including semantic classes, a dependent conjunction and a molecular property constraint; and guidance-scale sweeps that compare the step rules along their whole curves. These address the concerns about novelty, baseline coverage and generality. This comment states the contribution and summarises the results; the individual responses give the details.

Contribution

The paper studies guidance of a frozen diffusion model toward a hard constraint that is observable only through a binary check on finished samples. Its principal finding is that the magnitude of the conditional guidance, and in particular how it is distributed across samples, is consequential in its own right: with the guidance direction at each state preserved and the reference mean-step schedule matched at every timestep, replacing that distribution by equal steps lowers constraint satisfaction by 17.0 to 18.5 points at comparable FID_c (Section 6.2). The effect now holds in 32 configurations and with three estimator families, including the independently developed estimator of Spectral Guidance, which shows that the allocation effect is not specific to our trained estimator and persists across independently developed estimators. The theory of the paper places this finding on a foundation for hard constraints (the characterization of h for a discontinuous indicator, the contraction of the constraint signal with noise, and the relation between data-law and model-law targets), and the applications show its practical use: binary outcomes alone raise a molecular property filter from 54.3% to 96.8% and make guidance 53 to 55 times cheaper per satisfying sample than rejection at a 1.7% base rate.

Broader relevance

Methods that learn a conditional probability or value and then normalize, clip or adapt its gradient can use the matched-budget comparison of Section 6.2 to measure what that choice costs, separately from the overall guidance strength. The paper establishes the effect in the settings listed below; whether a given method should change its step rule is a question the comparison lets its authors answer.

1. Matched-budget experiment (Section 6.2)

The test preserves the guidance direction, matches the mean step at each timestep to that of the standard rule, and changes only how the step is allocated across samples (standard rule; equal step; 1/ĥ-weighted step).

The extensions show the effect in two demanding settings. On a negatively dependent conjunction, the equal step lowers joint satisfaction from 83.4 ± 0.3% to 76.1 ± 0.4% at γ = 10, with FID_c nearly unchanged (131.1 against 131.5). On the QM9 molecular property filter, it lowers satisfaction from 93.7 ± 0.4% to 91.0 ± 0.7% at γ = 5, while validity stays at 100% and uniqueness and novelty stay within one point. Both are reproduced over three sampling seeds. They take the allocation finding beyond individual image predicates and into a different generative domain.

The full list follows. A configuration is one estimator, one constraint, one guidance scale and one sampler. Thirty are on CIFAR-10 and two on QM9, with 1,000 samples per run.

SetEstimator (trained networks)ConstraintsScalesSampler; sampling seedsConfigurationsLoss of the equal step (points)
Table 2 of the paperclassifier (3 training seeds)red-pixel, upper luminanceγ = 21000 steps; seed 0 (seeds 0, 1 for classifier 1)617.0–18.5
Further scalesclassifier (1)red-pixel, upper luminanceγ = 1, 51000 steps; seed 048.4–22.4
Further predicatesclassifier (1)edge density, low saturationγ = 21000 steps; seed 0218.5–22.2
Coarser samplersclassifier (1)red-pixel, upper luminanceγ = 2100 and 50 steps; seed 046.2–17.4
Our head h_θh_θ (1)red-pixel, upper luminanceγ = 21000 steps; seed 0217.8–18.0
Class constraintsclassifier (1)cat, shipγ = 2, 5, 101000 steps; seed 0 (seeds 0, 1 at γ = 5)61.2–8.5
Spectral Guidance estimatortheir network (1)red-pixel, upper luminancestrength 0.25, 0.5DDIM 100 steps; seeds 0, 1, 241.8–5.9
Dependent conjunctiontwo composed heads h_θred-pixel AND low saturationγ = 5, 101000 steps; seeds 0, 1, 225.2–7.3
Molecular filter (QM9)our head h_θ (1)Lipinski-type filterγ = 2, 51000 steps; seeds 0, 1, 221.1–2.7

The table is discussed in part 2 of this response.

AuthorsOpenReview ↗

General response (2/2): discussion of the table, further results and summary

The table in part 1 covers eight constraints: four pixel predicates, two class labels, one conjunction and one molecular property filter. Table 2 uses three independently trained classifiers. The other experiments use one trained estimator per constraint, with the dependent conjunction combining two separately trained heads. The sampling seeds of each set are given in the table.

The equal step has lower satisfaction than the standard rule in all 32. The losses are 6.2 to 22.4 points on the pixel predicates with classifier estimators and our head, and 7.2 to 8.5 points on ship at γ = 5 and 10 (non-overlapping 95% intervals). They are 1.2 to 4.4 points on cat and 1.8 to 5.9 points with the Spectral Guidance estimator, where the difference exceeds the seed standard deviation in three of four configurations.

On the conjunction the equal step loses 5.2 points at γ = 5 and 7.3 points at γ = 10; at the latter scale all three seeds show losses between 7.0 and 7.6 points, with FID_c differing by less than 0.5. On the molecular filter it loses 2.7 points on average at γ = 5, between 1.5 and 3.3 points across the three seeds, and 1.1 points at γ = 2, with comparable validity, uniqueness and novelty.

The configurations cover pixel predicates, class labels and a conjunction on CIFAR-10, and one programmatic molecular property constraint on QM9. The allocation finding rests on these controlled experiments; the Stable Diffusion study of Section 6.7 addresses a different question, the feasibility of binary supervision on a latent text-to-image model.

2. Comparison with Spectral Guidance

Compared with the authors' released code, in their sampler, on the two predicates of Sections 6.1 and 6.2. Each method's strength is selected on sampling seed 0 and the comparison is reported on seeds 1 and 2: at high satisfaction (99.5 to 99.95%), our FID_c is 51.3 against 76.1 on red-pixel and 44.0 against 49.4 on upper luminance, with 35% less sampling time per sample. Protocol, tuning, timing and the relation between Proposition 2 and their spectral analysis are in the response to Reviewer 5LBY.

3. Normalized rule at smaller guidance scales

On CIFAR-10 (γ down to 0.1), each tested point of the normalized rule's curve has a point of ours with satisfaction at least as high and lower FID_c. On Stable Diffusion 1.5 we ran the scales below γ = 10 that Reviewer 5LBY requested; they do not improve on the normalized rule's γ = 10 point. The sweeps of both rules from γ = 10 upward are in Table 20 of the paper.

4. Composition of dependent constraints

For a pair that co-occurs in 1.28% of the data against 6.69% under independence, composing the two separately trained heads raises joint satisfaction from 0.5% to 83.4 ± 0.3% (3 seeds), and the matched-budget effect holds on this conjunction as well (table in part 1; response to Reviewer zHSD). This is one pair; composition is not guaranteed for arbitrary constraints.

5. Related work and presentation

We will cite and discuss Spectral Guidance, SAG, PAG, CFG++, Flow-GRPO and DiffusionNFT in the related-work section, state the relation between Proposition 2 and the spectral analysis of Spectral Guidance, and open the introduction by stating that the setting is conditional guidance with a binary label and that a noisy-state classifier estimates the same quantity. All results reported in these comments will be included in the paper.

Summary

Taken together, these additions address the principal concerns raised in the reviews. The comparison with Spectral Guidance establishes competitiveness against a recent, independently developed method. The extended controlled study establishes that the allocation effect persists across estimators, constraint types, sampler discretizations and generative domains. The theoretical analysis characterizes the hard-constraint conditional correction on which these experiments rest. The contribution is not a new expression for classifier guidance. It is a controlled empirical characterization of how the allocation of a known conditional correction affects hard-constraint satisfaction, with effects replicated across independent estimators, compound constraints and a molecular generation task. The matched-budget protocol makes this effect measurable separately from overall guidance strength and gives later guidance methods a concrete evaluation tool. Together with the hard-constraint analysis and the ability to steer frozen generators from binary evaluations of finished samples, these results form a broader contribution than the estimator alone.

AuthorsOpenReview ↗

Revised paper: summary of changes

We thank the reviewers for their constructive feedback and the thoughtful discussion, which helped us strengthen the presentation, theoretical positioning, and experimental evaluation of our work. We have uploaded a revised manuscript incorporating the changes discussed during the rebuttal period. For ease of reference, we summarize the main revisions below.

1. Introduction and contribution (Section 1). We have revised the introduction to establish the problem setting before presenting the mathematical formulation. We now explicitly clarify the relationship to classifier guidance with binary labels and distinguish our central contribution: the controlled characterization of how guidance-magnitude allocation affects constraint satisfaction.

2. Spectral Guidance comparison (Section 3 and Appendix D.1). We have incorporated the direct empirical comparison with Spectral Guidance, including the matched-budget experiment using its independently trained estimator. We also added the theoretical relationship between Proposition 2 and its spectral analysis, clarifying the corresponding results and assumptions.

3. Guidance-scale experiments (Section 6.7 and Appendix D.3). We have included the extended CIFAR-10 guidance-scale sweeps down to γ = 0.1 and the additional Stable Diffusion 1.5 experiments below γ = 10. Section 6.7 also clarifies why the independently tuned Stable Diffusion comparison does not constitute a controlled test of guidance-magnitude allocation.

4. Extended matched-budget evaluation (Section 6.2 and Appendices D.2, D.4, and D.5). We have incorporated the expanded study covering 32 configurations across three estimator families, including semantic class constraints, a strongly dependent conjunction, and a molecular property constraint. These experiments further evaluate the generality of the guidance-allocation finding beyond the original image predicates.

5. Expanded related work (Section 7). We have added and clarified the relationships with Spectral Guidance, SAG, PAG, CFG++, reward-based fine-tuning methods (Flow-GRPO, DiffusionNFT, and ReFL), Noise Hypernetworks, and distribution-matching distillation.

We sincerely appreciate the reviewers' detailed suggestions, which motivated these additions and helped us clarify both the scope and broader relevance of our contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.