acceptodds
Under review as a conference paper at ICLR 2027

Posterior Sampling by Source-Space MCMC via Prior-Based Few-Step Transport Maps

Abstract

Bayesian inference increasingly uses informative but implicit priors represented only by samples, such as historical ensembles, simulator outputs, and pretrained generative models. The same computational problem appears in the test-time guidance task (generalized Bayes), where an explicit positive weight, e.g., an exponentiated reward, tilts an implicit prior. We develop a framework for source-space generalized Bayesian inference that combines inexpensive few-step prior transports with posterior stability guarantees. Specifically, we represent the prior using a one- or few-step improved MeanFlow (iMF) map and perform posterior sampling in its Gaussian source space. We establish Wasserstein error bounds between the exact and learned posteriors in terms of the joint population iMF and auxiliary-velocity loss, decomposed into training suboptimality and model-class approximation error. In the iMF source space, we adopt parallel tempering with preconditioned Crank–Nicolson updates and introduce a hybrid variant that incorporates split Hamiltonian Monte Carlo to improve sampling efficiency. Synthetic experiments show that the proposed framework can approximate posterior distributions accurately and efficiently, while CLIP-guided ImageNet experiments demonstrate its ability to steer a pretrained iMF image prior toward text-specified preferences.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Hamiltonian Monte Carlo, posterior sampling, improved MeanFlow, parallel tempering, posterior stability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3j9rOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an efficient Bayesian posterior sampling framework using an implicit generative prior. Specifically, the authors use the existing improved MeanFlow method to learn a one-step or few-step transport map from a Gaussian source space to the data space, then perform posterior sampling in the source space using parallel tempering, preconditioned Crank-Nicolson, and hybrid HMC. Unlike existing flow matching approaches, the framework provides theoretical guarantees for the fixed few-step transport map used in practice, without discretization error from numerical ODE integration. Experiments on a synthetic Bayesian inference problem demonstrate higher sampling efficiency and posterior approximation accuracy than existing methods. The proposed method also achieves high rewards at low computational cost in CLIP-guided image generation.

Critical strengths

S1. This paper provides a connection between the population training objective of iMF, fixed few-step transport approximation error, and posterior stability. This addresses a limitation of prior results that consider exactly integrated flow maps.

S2. The proposed framework substantially reduces the cost of source-space MCMC compared with multi-step flow matching approaches, while maintaining competitive posterior approximation accuracy on synthetic experiments.

S3. The proposed method accommodates both conventional Bayesian likelihoods and reward-based test-time guidance without retraining the prior.

Critical weaknesses

W1. The synthetic experiments use a block-separable 32-dimensional posterior, while the high-dimensional ImageNet experiments focus only on reward-guided generation. This leaves the effectiveness of the proposed method on high-dimensional, strongly coupled Bayesian inverse problems largely untested. Given that the proposed framework is motivated by Bayesian inference with implicit generative priors, evaluating it on more realistic inverse problems would substantially strengthen the empirical validation. For instance, InverseBench (Zheng et al., ICLR 2025) provides challenging scientific inverse problems, including medical imaging and tomography, while A Statistical Benchmark for Diffusion Posterior Sampling Algorithms (Zach et al., ICLR 2026) provides reference posterior samples for quantitative evaluation of posterior accuracy and coverage. Experiments on such benchmarks would help establish whether the proposed method generalizes beyond simplified synthetic settings.

W2. The main components, including improved MeanFlow (iMF), source parallel tempering, preconditioned Crank-Nicolson, and split HMC, are already established methods. The algorithmic contribution consists primarily of their combination, which somewhat limits the novelty. The authors should more clearly articulate the novel aspects of the proposed framework.

W3. The Wasserstein bounds depend on potentially large, dimension- and model-dependent regularity constants, and the asymptotic δ1/8\delta^{1/8} dependence is weak. The applicability of the assumptions, particularly the positive lower bound on likelihood weights required for geometric convergence, is insufficiently demonstrated for practical Bayesian inverse problems. These assumptions require clearer justification.

W4. This paper does not systematically evaluate or ablate the influence of transport step count, auxiliary loss weighting, tempering configurations, or hybrid HMC hyperparameters on accuracy and computational efficiency. Such analyses are needed to clarify the novelty of the proposed framework and better substantiate its claims.

Questions for authors

Please see the weakness section.

3 comments

We thank the Reviewer for the thoughtful and constructive comments.

W1)

Thank you for the suggestion. We have added a 64-dimensional deconvolution problem with a coupled Laplace random-walk prior in Section 4.2, inspired by Zach et al. (2026). This is now a truly coupled posterior sampling problem whose posterior reference samples can be obtained via the Gibbs sampler thanks to the problem's structure. Evaluated across four observations, our hybrid sampler produces bulk effective samples 4.67 times faster than the pure pCN method using the same iMF transport. It also cuts sampling time by roughly two-thirds compared to the FM10 baseline while keeping distributional errors competitive.

References:

Martin Zach, Youssef Haouchat, and Michael Unser. A statistical benchmark for diffusion-posterior- sampling algorithms. In International Conference on Learning Representations, volume 2026, pp. 108984–109021, 2026.

W2)

We agree that the individual components are established. We already mentioned all of them in the introduction. Our contribution goes beyond combining them: we identify and analyze a formulation that connects the population training objective to the implemented few-step transport, posterior stability, and sampling error. The choice of transport objective is essential to this connection. For example, simply replacing iMF with few-step MF would not preserve our argument. Thus, few-step generation alone is insufficient for this guarantee. Please see also our reply to Reviewer RnYN (Weakness 2). We further extend the sampler convergence analysis to the hybrid SPT scheme and combine it with transport and posterior stability bounds to obtain an end-to-end error guarantee.

W3)

The assumption on the positive lower bound on likelihood weights (Assumption A.1) is used to establish geometric convergence of the source sampler; it is not required for our main results linking the population training objective to posterior approximation error. We note that the positive lower bound holds for bounded-reward guidance, Gaussian observation models with bounded forward-map range, and contamination models with a positive observation component independent of the unknown (see Appendix A.7). On the other hand, it fails for Gaussian likelihoods when the forward map is unbounded on the transport’s range. Our convergence theorem therefore does not cover that case.

W4)

We have added systematic ablations on the iMF+SPT+hybrid in Appendix B.4, Figures 9-14, covering all four requested aspects on the bana, sine and coupled Laplace problems. We vary one factor at a time and report sliced Wasserstein error and mean coordinate bulk ESS per second.

  • Transport steps: We test K∈{1,2,4,6,12}K\in\{1,2,4,6,12\}. More steps improve accuracy on sine and Laplace but reduce ESS per second. On banana, accuracy improves from one to two steps, with no consistent further gains.

  • Auxiliary loss weighting: We test λ∈{0,0.1,1,10}\lambda\in\{0,0.1,1,10\}. The lowest mean error occurs at λ=1\lambda=1 on Laplace and λ=0.1\lambda=0.1 on banana. Removing auxiliary supervision does not reduce accuracy on sine.

  • Tempering: We test T∈{1,8,16,24,48}T\in\{1,8,16,24,48\} temperatures and grid powers p∈{1,2,3}p\in\{1,2,3\}. Banana benefits in mean accuracy from moving from one to eight temperatures, but larger ladders offer no consistent gains. Sine and Laplace show little accuracy change. Grid power has little effect across all three problems.

  • Hybrid HMC: We test integration lengths L∈{2,6,12}L\in\{2,6,12\} and target acceptance rates {0.6,0.75,0.9}\{0.6,0.75,0.9\}. Increasing LL from 6 to 12 substantially improves ESS per second with similar accuracy across all three problems. Target acceptance affects efficiency more than accuracy.

I thank the authors for the rebuttal.

W1. The 64-dimensional deconvolution experiment partially addresses my concern that the original synthetic posterior was block-separable. I appreciate the inclusion of a coupled posterior and reference samples from Gibbs sampling. However, the evaluation still leaves open whether the method can accurately characterize the posterior and its uncertainty in realistic, high-dimensional inverse problems, such as medical imaging or tomography. This remains my main empirical concern.

W2. I appreciate the clarification that the main contribution is the theoretical connection between the iMF training objective, the implemented few-step transport, posterior stability, and sampling error.

W3. The authors clarify that the positive lower bound on likelihood weights is required for the sampler's geometric convergence result, while the main posterior approximation result does not rely on it. I appreciate this distinction. However, the response addresses only one part of my concern. I remain uncertain about how informative the Wasserstein bounds are in practice, given the potentially large dimension- and model-dependent constants and the weak asymptotic dependence on δ1/8\delta^{1/8}.

W4. The new ablations address all aspects I raised. I consider W4 resolved.

Overall, I appreciate the work that went into the rebuttal. Nevertheless, I was slightly disappointed that the authors did not directly address my central concerns. In particular, the responses to W1 and W3 felt evasive. The new experiment in W1 response leaves the method's performance on realistic, high-dimensional inverse problems untested, and the clarification about the convergence assumption leaves unanswered. Hence, I decided to lower my score to 1.

Thank you for your further engagement.

About the second part of W1, we cannot add large-scale medical imaging or tomography experiments within this rebuttal, due to time constraint. About the second part of W3, yes, the conservative constants and the rate δ1/8\delta^{1/8} are not ideal. It is just a theoretical guarantee, which can cover the worst case. Further justifying the usefulness of this bound in practice is left for furture work.

Reviewer_RnYNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper suggests learning a flow map for posterior sampling using a generative model as the prior. Improved MeanFlow (iMF) learns a few-step prior transport that maps from source space (z) to target space (x), and then posterior sampling can be performed using parallel tempering in source space (z). The paper theoretically proves that the posterior approximation error can be upper-bounded by the iMF training error. The authors demonstrate the efficiency and performance of their method on synthetic data and CLIP-guided ImageNet generation.

Critical strengths
  1. The paper provides a theoretical connection between the iMF training objective and the posterior approximation error for few-step transport maps.
  2. The proposed method significantly reduces the computational cost of posterior sampling by using few-step transport maps. In particular, the synthetic experiments demonstrate improvements in sampling efficiency while maintaining or improving posterior approximation accuracy.
Critical weaknesses
  1. The representative use cases of this method would be inverse problems and simulations, as mentioned in the paper. However, the experiments only include synthetic inverse problems and CLIP-guided ImageNet generation. Although the paper demonstrates improvements in sampling efficiency over existing baselines, it is unclear whether these improvements would also hold for more challenging applications involving complex posterior distributions and expensive simulations. Since the paper emphasizes the efficiency and practical applicability of the proposed method, the current list of tasks may not fully convince readers of the necessity of this method. I would recommend including at least one realistic inverse problem or scientific simulation task to demonstrate its practical advantages. We can refer to the experiments conducted in the previous work like Consistency Posterior Sampling (Purohit et al., 2025).

  2. The core formulation of the method seems similar to Consistency Posterior Sampling in terms of source-space posterior sampling with MCMC using a few-step generative model. Although the proposed method uses a different transport model (iMF) and sampling algorithms (SPT with pCN or hybrid HMC), the underlying approach of sampling from a Gaussian source-space posterior has already been explored in this prior work. Furthermore, Consistency Posterior Sampling also uses a few-step generative model, which makes it a particularly relevant baseline for evaluating the sampling efficiency of the proposed method. However, the paper only compares with diffusion posterior sampling and flow-matching-based methods, making it difficult to understand how much improvement the proposed method provides over existing few-step source-space posterior sampling methods. I strongly recommend comparing the proposed method with Consistency Posterior Sampling in terms of sampling accuracy and computational efficiency under the same computational budget. This comparison would better clarify the empirical contribution of the proposed method beyond its theoretical guarantees.

Questions for authors

What is the critical contribution over Consistency Posterior Sampling? I think the formulation is similar but the detailed method variants are different (iMF vs. Consistency Model, langevin vs. parallel tempering. Is there any fundamental advantage that overcomes an issue of Consistency Posterior Sampling? This could be helpful to understand the novelty of this paper.

3 comments

We thank the Reviewer for the thoughtful and constructive comments on our work.

Weakness 1

As also suggested by Reviewer 3j9r, in the revision, we added a 64-dimensional deconvolution problem with a coupled Laplace random-walk prior, inspired by Zach et al. (2026); see Section 4.2 in the revised manuscript. Across four observations, the hybrid achieves 4.67 times higher bulk ESS per second than all-pCN using the same iMF transport. Compared with the FM10 baseline, it requires approximately three times less sampling time, with competitive distributional errors. These results provide additional evidence for coupled inverse problems.

References:

Martin Zach, Youssef Haouchat, and Michael Unser. A statistical benchmark for diffusion-posterior- sampling algorithms. In International Conference on Learning Representations, volume 2026, pp. 108984–109021, 2026.

Weakness 2

We agree that Consistency Posterior Sampling (CPS) is closely related. Consistency models, MeanFlow (MF), and improved MF (iMF) all enable few-step transports, and we do not claim that few-step source-space posterior sampling is itself novel.

Our central contribution is establishing how the population few-step training objective controls the resulting posterior error, which is important towards reliable posterior sampling. This connection, however, does not follow automatically from few-step generation. As discussed in the introduction and Section 2.2., for the original MF, the parameter-dependent gap between the training loss and marginal risk prevents such a guarantee (as far as we are concerned). CPS, on the other hand, already provides posterior stability guarantees assuming prior approximation accuracy, but does not establish how its consistency-model training objective controls the resulting posterior error, as discussed in our introduction section. Our contribution therefore goes beyond substituting one few-step model for another: we identify and analyze a formulation that enables control of posterior error through the population training objective.

Algorithmically, we use source-space parallel tempering with Metropolis-corrected pCN or split HMC aiming at high-dimensional, multimodal regimes. We agree that a matched-budget comparison with CPS’s Langevin sampler provides more practical insight. However, we believe a direct comparison is less informative because it changes both the learned prior and the source-space sampler simultaneously.

Question: This is answered in our reply to Weakness 2.

I acknowledge that connecting the population iMF training objective to posterior approximation error is a meaningful theoretical extension over CPS. However, CPS already provides posterior stability guarantees based on prior approximation accuracy. Therefore, the key remaining question is whether the additional theoretical guarantees translate into practical improvements in posterior approximation accuracy or computational efficiency.

I understand that a direct comparison with CPS changes both the learned prior and the sampler. Nevertheless, comparing the complete frameworks under the same computational budget is necessary to evaluate their practical advantages. A controlled comparison using the same iMF prior with CPS-style Langevin sampling would further help isolate the contribution of the proposed sampler.

We thank the Reviewer for the reply. We are glad that the Reviewer recognizes our theoretical contribution as meaningful.

We now included the CPS framework for the ImageNet experiment in the updated manuscript (see Section 4.3, Table 3, and Appendix B.5). We use the modern consistency model FACM [1], with the pretrained checkpoint 400ep-stg2.pt (https://huggingface.co/Peterande/FACM/blob/main/400ep-stg2.pt for 256x256 ImageNet) and perform Unadjusted Langevin Algorithm (ULA) in its source space.

The results show a reward-cost trade-off: with the same MCMC length, FACM+ULA is faster, whereas our methods achieve higher CLIP rewards. Within-chain CLIP diversity varies across class-prompt pairs, neither methods dominate. Visual inspection also reveals a difference in the sampled trajectories: FACM+ULA typically produces gradual changes between successive images, whereas accepted replica exchanges in our PT methods allow abrupt transitions between distinct image configurations. This provides qualitative evidence that replica exchange contributes to exploration beyond local updates. See an instance (Class: bald eagle, prompt: “A photograph of a bald eagle soaring above snow-capped mountains, wings spread wide.”) in https://anonymous.4open.science/r/Additional-results-C654/README.md

Nevertheless, we include FACM+ULA primarily as a reference for a modern CM-based CPS implementation.

Reference

[1] Peng, Yansong, Kai Zhu, Yu Liu, Pingyu Wu, Hebei Li, Xiaoyan Sun, and Feng Wu. "Facm: Flow-anchored consistency models." In International Conference on Learning Representations, vol. 2026, pp. 7520-7543. 2026.

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

The paper considers Bayesian and generalized Bayesian inference with implicit priors represented by samples. It uses iMF to construct few-step prior transports and performs posterior sampling in the Gaussian source space through parallel tempering with pCN updates. A hybrid variant replaces the coldest replica’s update with split HMC. The theoretical analysis connects the joint population training risk to transport error and posterior Wasserstein error under regularity assumptions. Experiments on synthetic distributions and CLIP-guided ImageNet generation evaluate posterior approximation, sampling efficiency, and reward improvement.

Critical strengths

The analysis addresses the implemented few-step transport maps. This is a useful contribution because a guarantee for an exactly integrated learned flow does not by itself establish the accuracy of a fixed few-step implementation. The main arguments appear sound under the stated assumptions, and the extension of the convergence analysis to the hybrid sampler provides additional theoretical support.

Critical weaknesses
  1. The applicability of the theoretical guarantees to the evaluated models remains insufficiently established. The population objective analyzed in Section 3.2 differs from the reported training recipe of the pretrained ImageNet iMF model in time sampling, loss weighting, and classifier-free guidance. The posterior stability bound also appears extremely conservative for the synthetic examples, while the positive lower bound required by the sampler convergence theorem is not established for their Gaussian likelihoods. These limitations matter because the connection between training and posterior accuracy is a central contribution. The paper should distinguish more clearly between the conditional mathematical results and the guarantees available for the actual experiments.

  2. The experiments do not yet establish the accuracy–cost advantage of the proposed approach. The runtime comparisons use six iMF evaluations against 400 FM evaluations in the synthetic setting, and one iMF step against 125 Heun steps for SiT. They demonstrate savings under these configurations, but leave the performance of cheaper FM and SiT configurations unresolved. The image experiments additionally compare different pretrained priors and primarily evaluate the same CLIP reward used to define the target. Without sampling diagnostics or independent quality and diversity measurements, the results do not establish whether the retained samples adequately explore that target. This is particularly relevant to the hybrid sampler, whose image-sampling cost is approximately 4.5 times that of all-pCN without a clear improvement in the reported outcomes.

Questions for authors
  1. Which guarantees are intended to apply to the pretrained ImageNet model? Section 3.2 assumes independent uniform time sampling and an unweighted joint squared loss, whereas the reported iMF training recipe uses logit-normal sampling, samples with r=tr=t, adaptive weighting, and classifier-free guidance targets. Can the authors explain how the actual objective controls the population risk in Theorem 3.7?

  2. How informative is Theorem 3.9 for the stated synthetic problems? Evaluating its condition LyE≤Zy/2L_y\sqrt{\mathcal E}\le Z_y/2 gives RMS transport-error thresholds of approximately 2.22×10−162.22\times10^{-16} for Banana and 1.40×10−161.40\times10^{-16} for Sine. These are sufficient conditions, so exceeding them does not imply inaccurate sampling, but they appear too restrictive to assess the reported results. Can the authors provide an informative instantiation of the bound and clarify whether the Gaussian-likelihood experiments satisfy the assumptions of the sampler convergence result?

  3. Does the computational advantage remain when FM and SiT use smaller integration budgets? Accuracy and runtime measurements across several budgets would help establish the cost of reaching comparable performance. For ImageNet, please also clarify how differences in prior quality and guidance settings affect the comparison.

  4. Do the retained image chains provide evidence of adequate mixing? Please report reward or feature ESS, agreement across independent chains, and replica-exchange diagnostics. An independent assessment of image quality and diversity would also help interpret the CLIP results. In particular, does the hybrid sampler improve ESS per second enough to justify its additional gradient cost?

3 comments

We thank the reviewer for the thoughtful and constructive comments. Because the weaknesses and questions overlapped, we addressed them together.

W1: The population objective analyzed in Section 3.2 differs... Q1: Which guarantees are intended to apply to the pretrained ImageNet model?...

Our analysis focuses on the plain iMF formulation, which captures the mathematical core connection between training and posterior error. This plain training setting is used in our synthetic experiments. The pretrained ImageNet model of Geng et al. (2026) uses additional empirically motivated recipes for large-scale image training, such as adaptive loss weighting and classifier-free guidance. These modifications are not well-analyzed theoretically in general and outside our current analysis and are not central to our contribution. Some extensions, such as alternative time-sampling distributions, are possible. We clarify this scope in Section 4.3 of the revised manuscript.

References

Geng et al. Improved Mean Flows: On the Challenges of Fastforward Generative Models. CVPR, 2026.

W1: The posterior stability bound also appears extremely conservative for the synthetic examples. Q2: How informative is Theorem 3.9...? Evaluating its condition LyE≤Zy/2L_y\sqrt{\mathcal E}\le Z_y/2 gives...

As the reviewer remarks, LyE≤Zy/2L_y\sqrt{\mathcal E}\le Z_y/2 in Theorem 3.9 is a sufficient condition, not a necessary condition. We agree that the global estimates for Zy/LyZ_y/L_y can be conservative (small) for the Gaussian synthetic examples, hence requiring a more accurate transport approximation E:=EZ∼γ∥Sθ(Z)−S(Z)∥2\mathcal{E}:=\mathbb E_{Z\sim\gamma}\|S_\theta(Z)-S(Z)\|^2. However, the Zy/LyZ_y/L_y can also be moderate for other problems. For example, consider the logistic likelihood ℓy(x)=(1+e−ya⊤x)−1\ell_y(x)=(1+e^{-ya^\top x})^{-1}, where y∈{−1,1}y\in\{-1,1\}, ∥a∥=1\|a\|=1, and the prior is symmetric about the origin, i.e., X=d−XX\overset{d}{=}-X. Since ℓy(x)+ℓy(−x)=1\ell_y(x)+\ell_y(-x)=1, we have E(ℓy(X))+E(ℓy(−X))=1E(\ell_y(X)) + E(\ell_y(-X))=1, and then Zy=E(ℓy(X))=1/2Z_y=E(\ell_y(X))=1/2. Moreover, ∥∇ℓy(x)∥=ℓy(x)(1−ℓy(x))≤1/4\|\nabla\ell_y(x)\|=\ell_y(x)(1-\ell_y(x))\le1/4, so Ly=1/4L_y=1/4. Thus Zy/Ly=2Z_y/L_y=2, and the sufficient condition becomes E≤1\sqrt{\mathcal E}\le1, independently of the ambient dimension.

Furthermore, the condition LyE≤Zy/2L_y\sqrt{\mathcal E}\le Z_y/2 can be removed for some models. Indeed, in Theorem 3.9, we use that condition only to ensure Zθ,yZ_{\theta,y} is bounded away from 00 for all θ\theta, as shown in the theorem's statement. Therefore, if the model already admits a positive lower bound on Zθ,yZ_{\theta,y}, we do not need that condition. Three examples are (see Appendix A.7 in the revised version):

  • Bounded-reward guidance: ℓy(x)=eτRy(x)\ell_y(x)=e^{\tau R_y(x)} where Ry(x)≥rmin⁡R_y(x)\geq r_{\min} for every xx, it holds Zθ,y≥eτrmin⁡Z_{\theta,y} \geq e^{\tau r_{\min}}, ∀θ\forall \theta.

  • Contaminated observation models: for p(y∣x)=(1−ε)psig(y∣x)+εq(y)p(y\mid x)=(1-\varepsilon)p_{\mathrm{sig}}(y\mid x)+\varepsilon q(y), where qq is a fixed observation density independent of xx, 0<ε<10<\varepsilon<1, and q(y)>0q(y)>0, we have Zθ,y≥εq(y)Z_{\theta,y}\ge\varepsilon q(y), ∀θ\forall \theta.

  • Gaussian observation models with bounded forward-map range: consider y=G(x)+ηy=G(x)+\eta, where η∼N(0,σ2Idy)\eta\sim\mathcal N(0,\sigma^2I_{d_y}) with fixed σ>0\sigma>0. Suppose ∥G(x)∥≤B<∞\|G(x)\|\le B<\infty for every xx. Then, Zθ,y≥(2πσ2)−dy/2exp⁡(−(∥y∥+B)22σ2)Z_{\theta,y} \geq (2\pi\sigma^2)^{-d_y/2} \exp\left(-\frac{(\|y\|+B)^2}{2\sigma^2}\right)

W1: the positive lower bound required by the sampler convergence theorem is not established... Q2: clarify whether the Gaussian-likelihood experiments...

The uniform positive lower bound ℓy(x):=p(y∣x)≥m>0\ell_y(x):=p(y|x) \geq m>0 in Assumption A.1 has not been established for the Gaussian-likelihood case with linear observations. However, we only need Assumption A.1. for geometric convergence of the source sampler, but not for the main theorems connecting the prior training error to the posterior error (Theorems 3.7 and 3.9). We also note that Assumption A.1 holds for bounded-reward guidance models, Gaussian observation models whose forward maps have bounded range, etc. (see Appendix A.7).

W2: leave the performance of cheaper FM and SiT configurations unresolved. Q2: computational advantage remain when FM and SiT use smaller integration budgets?

iMF is designed natively for few-step generation, with theoretical control of the implemented transport under our assumptions. FM and SiT rely on numerical integration of continuous flows, hence, choosing few steps cannot generally be assumed accurate without validation

We have expanded the comparison to include cheaper FM and SiT configurations.

For FM, Appendix B.3 evaluates 2, 5, 10, 50, and 100 RK4 steps, keeping trained models and other sampler settings fixed. Using 5 and 10 steps substantially improves the baseline’s accuracy-cost tradeoff. Nevertheless, across the three synthetic problems, our hybrid is 2.9-3.3 times faster and achieves 26-48 times higher bulk ESS per second than 10-step FM, with competitive distributional errors. See next comment

continue from the previous comment

For SiT, we additionally evaluate 10 and 50 Heun steps on eight class-prompt pairs. SiT10 improves reward over SiT50 on seven prompts, but both iMF variants outperform both configurations on all eight. Mean CLIP rewards are 0.3414 for SiT10, 0.3348 for SiT50, 0.3734 for iMF+pCN, and 0.3732 for iMF+hybrid. The observed reward advantage therefore persists against these cheaper SiT configurations. The detailed results are given below.

CLIP rewards

Class / promptSiT50SiT10iMF+pCNiMF+hybrid
Aircraft carrier / sunset0.35280.35950.36670.3659
Aircraft carrier / watercolor0.30960.31590.37360.3774
Bald eagle / autumn perched0.32090.32520.37430.3781
Bald eagle / fishing0.34990.35540.36250.3596
Bald eagle / mountain flight0.35520.35250.39360.3914
Bald eagle / snow painting0.32880.34000.37270.3713
Castle / snow painting0.29940.31990.36180.3569
Castle / sunrise0.36190.36310.38200.3849
Average0.33480.34140.37340.3732

Q3: For ImageNet, please also clarify how differences in prior quality and guidance settings affect the comparison.

We clarify that both pretrained generators use classifier-free guidance. Both models have strong published ImageNet results, with guided iMF-XL/2 reporting FID 1.72 and SiT-XL/2 reporting 2.15. Differences in prior quality and CFG can therefore contribute to the observed reward and diversity. Our ImageNet results establish the reward-cost performance of the evaluated pretrained configurations, rather than isolating the sampler’s contribution. In contrast, the comparison between our all-pCN and hybrid samplers uses exactly the same iMF prior and guidance settings.

W2: Without sampling diagnostics or independent quality and diversity measurements, the results do not establish... Q4: Do the retained image chains provide evidence of adequate mixing?

In the revised manuscript, we now report CLIP-feature diversity alongside reward, providing a broader assessment of the generated samples under the available computational budget. The results demonstrate that high reward can be achieved while retaining feature diversity. This supports the practical value of the framework for guided generation, although diversity alone does not establish convergence to the target distribution. It would be very demanding to do so because of the curse of dimensionality. The image-chain R^\widehat R diagnostics indicate that longer runs would be needed for convergence/mixing claims. However, our compute budget does not allow substantially longer runs.

The authors further clarified the conditions and scope under which the theory applies and provided examples to help me better understand it. In addition, they included experiments with cheaper FM and SiT configurations, which demonstrate the computational advantages of the method. I think my main concerns have been addressed, and I am willing to raise my score to 3.

Comments on the paper

1 comment

We thank the Reviewers for their time in reviewing our manuscript. We acknowledge that all the comments are accurate. We have updated our manuscript PDF, where changes are highlighted in blue.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.