acceptodds
Under review as a conference paper at ICLR 2027

MetaDMD: Stable and Scalable Training of Posterior Samplers for Test-Time Steering

Abstract

Test-time steering aligns pretrained diffusion and flow models with arbitrary rewards without fine-tuning, but it requires sampling clean data from the posterior given each intermediate noisy state. Exact posterior sampling via SDE or ODE simulation is computationally expensive, and repeating it at every step makes the cost prohibitive. Stochastic flow maps reduce this cost by learning a flow map that transports Gaussian noise to the posterior in a few steps. However, they rely on consistency objectives that distill the exact integration of this trajectory, which makes training unstable and slow to converge. We observe that steering requires only samples that follow the posterior, not the trajectory that reaches them. Based on this, we propose MetaDMD, which learns a few-step posterior sampler by matching posterior distributions instead of individual trajectories. MetaDMD requires no consistency loss, or pretrained flow map, thus trains stably even on large text-to-image (T2I) models. On class-conditional ImageNet with SiT-XL/2, MetaDMD outperforms stochastic flow maps in test-time steering with reward gradients and performs comparably without them. On SD3.5-M, it also outperforms them across various rewards on DrawBench and on GenEval, while converging approximately faster.

Primary area:
generative models
Keywords:
Diffusion Models, Flow Matching, Distillation, Test-time Steering

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_cdL2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes MetaDMD, a few-step posterior sampler for test-time steering. By extending distribution matching distillation to conditional posterior sampling, MetaDMD avoids explicit trajectory matching and provides a more efficient training framework for posterior samplers. The method achieves improvements over existing baselines on both ImageNet and text-to-image generation tasks.

Critical strengths
  1. By extending distribution matching distillation to conditional posterior sampling, this paper explores a well-motivated and theoretically grounded direction for efficient test-time steering.
  2. The evaluation on large-scale text-to-image models demonstrates that the proposed method is not limited to small-scale settings and provides convincing evidence of its practical effectiveness.
Critical weaknesses

The claim that avoiding trajectory matching is the main reason for faster convergence may be overstated. While MetaDMD avoids explicit trajectory-level losses, DMD still performs distribution matching along the probability path. Therefore, the improvement may come from the optimization behavior of the DMD framework itself, which is not sufficiently disentangled in the experiments.

Questions for authors

Why is DMD used instead of DMD2? Given that adversarial training has shown benefits for one-step and few-step generative distillation, could a DMD2-style objective further improve the quality and stability of MetaDMD?

5 comments

We thank the reviewer for the valuable feedback, mainly asking about the source of the faster convergence of MetaDMD and our choice to omit the adversarial loss of DMD2. All revisions are shown in skyblue in the revised PDF, and those addressing your comments are marked with cdL2 in orange.

W1. Source of faster convergence.

We agree that DMD matches distributions at multiple noise levels along the probability path, and we would like to clarify why this does not act as a trajectory constraint. In MetaDMD, the student output is noised with fresh Gaussian noise that is independent of the inner noise z0z_0 used to generate it. The noised distribution at every noise level is therefore determined only by the student’s output distribution p1∣tθ(⋅∣xt)p^\theta_{1|t}(\cdot \mid x_t), so matching it across noise levels is equivalent to matching this endpoint distribution. The trajectory distillation of MFM instead requires each z0z_0 to pass through the intermediate states zsz_s along the inner ODE path, which MFM enforces with consistency objectives that are known to be unstable to train [1, 2], and we believe that removing this requirement contributes to the efficient training of the sampler (Table (a)). We have clarified this point in the revised manuscript (Sec. 3.1).

Table (a) GenEval score on SD3.5-M, N=8, gradient-free estimator. (Fig. 3, Sec. 4.2; Tab. 11, App. F)

MethodTraining iterationsCumulative GPU-hoursGenEval score
Unsteered teacher baseline——0.672
MetaDMD00.000.688
MetaDMD1,0008.380.776
MetaDMD2,00016.750.776
MFM (Teacher initialization)00.000.556
MFM (Teacher initialization)1,00011.240.691
MFM (Teacher initialization)2,00022.470.707
MFM (Teacher initialization)3,00033.710.754
MFM (FlowMap initialization)010.220.589
MFM (FlowMap initialization)1,00013.790.661
MFM (FlowMap initialization)2,00017.370.672
MFM (FlowMap initialization)10,00045.970.714

Q1. DMD versus DMD2.

We thank the reviewer for pointing out the possibility of applying adversarial loss to MetaDMD. Distribution matching remains a core component of DMD2, and our objective is to extend it to conditional posterior sampling. We largely follow DMD2’s implementation, including the two-time-scale update rule, but omit the adversarial loss to isolate the contribution of the distribution-matching objective for training the stochastic posterior sampler. We have clarified these implementation choices and their relationship to DMD2 in the revised manuscript (App. D), and leave the integration of adversarial loss to MetaDMD as a future work.

[1] Hu et al. CMT: Mid-training for Efficient Learning of Consistency, Mean Flow, and Flow Map Models. In ICLR, 2026.

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

I am still not fully convinced by the explanation. Although independently re-noising the student output removes the explicit sample-wise correspondence with the teacher's ODE trajectory, isn't the gradient loss itself still exploiting the teacher's denoising-field information along the probability path through the difference between the real and fake denoisers? In this sense, teacher-specific denoising information is still incorporated into the supervision, even though explicit trajectory matching is removed. I would appreciate a more explicit clarification of this point.

We thank the reviewer for this follow-up question. We agree with the reviewer that MetaDMD also exploits the teacher's denoising-field information. However, the key difference lies in how this information is used. DMD uses the teacher's denoising-field information to compute the score difference between the teacher and student distributions, which provides the gradient for minimizing the KL divergence between two distributions in the noised space. In contrast, flow maps use the same denoising-field information in consistency objectives to approximate the exact integration of the teacher's ODE, i.e., trajectory. This strong constraint leads to unstable, costly optimization behavior. MetaDMD avoids such trajectory-level consistency objectives, which we believe contributes to its more efficient training.

In summary, both DMD and flow maps use the teacher's denoising-field information but impose different constraints: DMD matches the student and teacher distributions, whereas flow maps aim to reproduce the exact individual trajectories induced by the teacher's ODE.

Reviewer_qA8BOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper uses DMD to train a meta flow map to match the transition kernel p1∣t(⋅∣xt)p_{1|t}(\cdot\mid x_t) of an underlying diffusion (or flow) SDE and then does guidance experiments on t2i and class-conditional experiments on imagenet comparing to vanilla meta flow maps.

Critical strengths
  1. The paper is well-written and presents an interesting idea of using GAN-style DMD loss on the transition kernel for the diamond/meta flow map type model of the GLASS conditional ODE.
  2. Simple idea and a natural extension of MFMs / diamond maps.
  3. The training efficiency argument seems compelling in §4.3.
Critical weaknesses
  1. I don't agree with §4.2, both MFM [1] and diamond maps [2] can be trained via a DMF-styled [3] distillation which can be very cheap to do in practice, ie, fine-tuning from a teacher. This doesn't seem to be a very fair comparison. This is done in practice by MFM in [1, §7.2.1] and could be trivially done with diamond maps. This is mentioned later in the imagenet comparisons. The results also seem to differ from the stated results in [1, Figure 7].
  2. The paper only compares to MFM in it's experiments and as mentioned above in some settings which I am not convinced are fair. In particular the paper does not compare to Poster Diamond and more importantly Weighted diamond maps which are training-free w.r.t. to standard off-the-shelf flow map which would alleviate the training concerns raised early.
  3. I'm very concerned that the gradient-based estimator underperforms the gradient-free estimator this is the inversion of the results [1] and other general results on gradient vs non-differentiable guidance in the literature. As such I'm fairly concerned about the reliability of the results. Also why are there no gradient-free results for Table 3? Also why is MFM worse than the teacher, but this is inverted for Table 3 with imagenet? This makes me think the model is poorly trained rather than metaDMD truly beating it.
  4. There is no comparison to other baselines like best-of-N search strategies for non-differentiable rewards or just standard ODE style guidance with full backprop, tweedie lookahead [4], or flow map lookahead [5].

References

[1] Peter Potaptchik, Adhi Saravanan, Abbas Mammadov, Alvaro Prat, Michael Samuel Albergo, and Yee Whye Teh (2026). Meta flow maps enable scalable reward alignment. arXiv (also published at ICML)

[2] Peter Holderrieth, Douglas Chen, Luca Eyring, Ishin Shah, Giri Anantharaman, Yutong He, Zeynep Akata, Tommi Jaakkola, Nicholas Matthew Boffi, and Max Simchowitz (2026). Diamond maps: Efficient reward alignment via stochastic flow maps. arXiv & ICML

[3] Lee, Kyungmin, Yu, Sihyun, and Shin, Jinwoo (2026). “Decoupled MeanFlow: Turning Flow Models into Flow Maps for Accelerated Sampling”. In: The Fourteenth International Conference on Learning Representations.

[4] Chung, H., Lee, S. and Ye, J.C., 2024, May. Decomposed diffusion sampler for accelerating large-scale inverse problems. In International conference on learning representations (Vol. 2024, pp. 38922-38949).

[5] Sabour, A., Albergo, M.S., Domingo-Enrich, C., Boffi, N.M., Fidler, S., Kreis, K. and Vanden-Eijnden, E., 2025. Test-time scaling of diffusions with flow maps. arXiv preprint arXiv:2511.22688.

Questions for authors
  1. Why is Table 1 comparing DMD vs MetaDMD, the more apt comparison is MFM vs MetaDMD no?
  2. Why do the gradient-free estimators outperform the gradient-based estimators? This seems to be in contrast to prior literature.
  3. Why do the authors posit that DMD style training of the flow map would be better than the conditional ODE distillation? As the condition ODE only matches in law (at the endpoint) and not in path the diffusion SDE. Is it because the midpoints on the "new" time axis don't match the SDE in law? Interestingly, an ICLR workshop paper came up with a different conditional ODE from GLASS which ensures that the intermediate time points equal the SDE in law and not just the terminal time point in the "new" time axis. Likewise, this paper seems to be moving in the opposite direction from other works on stochastic flow maps [2,3] which learns a stochastic flow map that is pathwise equivalent along the whole trajectory, whereas this paper seems to double down on learning only the weak (in law) flow map matching the distribution but not paths of either the true SDE (or GLASS-style) conditional ODE. I'm curious how the analysis carries out here with this DMD style of training the stochastic flow map.

I generally like the paper and think the core idea of the paper is good, but I'm fairly concerned about the results and comparisons to MFM and missing comparisons to diamond maps.

References

[1] Passaro, R., Blasingame, Z.W., Bronstein, M.M. and Tong, A., 2026. Stochastic Few-step Models.

[2] McCallum, S., Blasingame, Z.W., Herschell, T., Rindtorff, N., Tong, A. and Foster, J., 2026. Strong stochastic flow maps. arXiv preprint arXiv:2606.01086.

[2] Pan, Z., Potaptchik, P., Yao, W., Albergo, M.S. and Pidstrigach, J., 2026. It^ o maps for any-step SDEs. arXiv preprint arXiv:2606.11156.

7 comments

We thank the reviewer for the detailed and constructive review. The reviewer’s questions mainly ask about the fairness of our comparison with MFM and the reliability of our results, along with comparisons to additional baselines. All revisions are shown in skyblue in the revised PDF, and those addressing your comments are marked with qA8B in green.

W1-1. Comparison fairness.

We thank the reviewer for raising this point and would like to clarify the training setup of MFM. The setting in MFM §7.2.1 does not fine-tune directly from the flow-matching teacher but from the DMF-XL/2+ flow map. Their setting trains the flow map into MFM for 100K iterations with a batch size of 360, which takes approximately 1,050 GPU-hours on NVIDIA RTX A6000 GPUs. In comparison, MetaDMD is trained directly from the flow matching teacher SiT-XL/2 for 2K iterations with a batch size of 128, which takes approximately 5 GPU-hours on the same hardware (Sec. 4.3). Moreover, the DMF-XL/2+ flow map used for training MFM is itself distilled from SiT-XL/2 for 400K iterations, four times the number of iterations in the MFM training stage. We hope the reviewer agrees that this pretraining stage is also part of the training cost of MFM, which MetaDMD does not require.

We also report the same trend of training efficiency on SD3.5-M, in Table (a). Table (a) shows the GenEval score against cumulative training GPU-hours, and MetaDMD reaches a final score of 0.776 after 16.75 GPU-hours. MFM initialized from a flow map reaches only 0.714 even with more than 2.5 times the training compute. MFM initialized directly from the teacher, which corresponds to the fine-tuning setting suggested by the reviewer, also reaches only 0.754 even with twice the training compute of MetaDMD.

Table (a) GenEval score on SD3.5-M, N=8, gradient-free estimator. (Fig. 3, Sec. 4.2; Tab. 11, App. F)

MethodTraining iterationsCumulative GPU-hoursGenEval score
Unsteered teacher baseline——0.672
MetaDMD00.000.688
MetaDMD1,0008.380.776
MetaDMD2,00016.750.776
MFM (Teacher initialization)00.000.556
MFM (Teacher initialization)1,00011.240.691
MFM (Teacher initialization)2,00022.470.707
MFM (Teacher initialization)3,00033.710.754
MFM (FlowMap initialization)010.220.589
MFM (FlowMap initialization)1,00013.790.661
MFM (FlowMap initialization)2,00017.370.672
MFM (FlowMap initialization)10,00045.970.714

W1-2. Inconsistent results.

For all ImageNet experiments, we use the official checkpoint of MFM-XL/2, and our results with this checkpoint follow the same trend as Figure 7 of the MFM paper. In both cases, gradient-based MFM steering provides substantial reward improvements and outperforms the gradient-free estimator. Absolute scores are not directly comparable because Figure 7 uses 250 outer steps on a single fixed ImageNet class whereas our experiments use 28 outer steps averaged over 100 classes. We have clarified this evaluation protocol together with the use of the official checkpoint in the revised manuscript (Sec. 4.1; App. D).

W2-1. Weighted Diamond Maps comparison.

In our additional experiment on SD3.5-M, MetaDMD achieves a higher GenEval score than Weighted Diamond Maps (0.776 vs. 0.580) under matched inference compute budgets. For Weighted Diamond Maps, we used a publicly available SD3.5 flow-map checkpoint and followed the original paper’s choice of four surrogate reward functions and their linear-combination weights. We set the number of particles to 15 to match MetaDMD’s inference compute budget.

An important distinction is that Weighted Diamond Maps require differentiating the reward through the predicted clean image and the flow map, preventing it from directly using the non-differentiable GenEval reward. In contrast, the gradient-free estimators of MFM and MetaDMD can steer directly with GenEval. Its lower performance may also partly reflect the difficulty of gradient-based steering on SD3.5-M: from our experiments in the manuscript, gradient-based steering underperforms gradient-free steering for both MFM and MetaDMD on this model. We have added this comparison to the revised manuscript (Sec. 4.2; App. E.5).

W2-2. Posterior Diamond Maps comparison.

We used MFM as the representative baseline because MFM and Posterior Diamond Maps share essentially the same design. Both distill the GLASS Flows ODE into a few-step stochastic flow map using consistency-based trajectory distillation, and both are initialized from a pretrained flow-map (MeanFlow-type) checkpoint. We have clarified this relationship and our rationale for selecting MFM in the revised manuscript (Sec. 4.1).

W3-1, Q2. Inverted trend of using gradient

We thank the reviewer for highlighting this discrepancy. The relative performance of the two estimators changes between class-conditional SiT and text-conditioned SD3.5-M. This reversal occurs for both MFM and MetaDMD, which suggests that it reflects a property of SD3.5-M rather than an issue specific to either posterior sampler.

To examine this further, we additionally ran DPS [2] on SD3.5-M, a standard gradient-based guidance method. Since it does not use a trained posterior sampler, its behavior is independent of how MFM or MetaDMD is trained. We ran DPS with each of the four rewards on DrawBench using the same steering steps as MetaDMD. As shown in Table (b), DPS falls below the unsteered teacher on ImageReward, PickScore and HPSv2 and improves the aesthetic score only marginally. This suggests that reward gradients on SD3.5-M are difficult to exploit regardless of the posterior sampler, which is consistent with the weaker gradient-based results in Table 2.

We conjecture that architectural differences between SiT and SD3.5-M lead to different gradient behavior when the reward is backpropagated with respect to the noisy state. Because the gradient-based estimator relies on the derivatives of the sampler, it is more sensitive to this behavior than the gradient-free estimator, which only uses reward values to reweight posterior samples (Sec. 4.2; App. E.4).

Table (b) Gradient-based steering on SD3.5-M, DrawBench. (Tab. 4, Sec. 4.2)

MethodImageReward ↑PickScore ↑HPSv2 ↑Aesthetic ↑
Teacher (no steering)0.7922.270.2765.48
DPS0.5621.970.2705.54
MFM, gradient-based (N=8)0.7422.230.2745.50
MetaDMD, gradient-based (N=8)0.9922.290.2855.62
MetaDMD, gradient-free (N=8)1.3322.930.2975.83

W3-2. Gradient-free results for Table 3.

Thank you for pointing out the missing reference. The corresponding gradient-free results are reported in Table 6 of the appendix (Tab. 8, App. E of the revised manuscript), where they were included due to space constraints in the main paper. We have added an explicit reference alongside Table 3 to make these results easier to locate (Sec. 4.3).

W3-3. MFM underperforms the teacher.

Under the training budget reported in the manuscript, MFM indeed underperforms the unsteered teacher in some SD3.5-M settings. To examine whether this gap comes from insufficient training, we additionally trained MFM with more than 2.5 times the GPU hours used for MetaDMD from a flow-map initialization and with twice the GPU hours from a teacher initialization (Table (a); Fig. 3, Sec. 4.2). Even with these extended budgets, MFM reaches GenEval scores of only 0.714 and 0.754, respectively, which are still below MetaDMD’s 0.776. This suggests that the gap does not stem from insufficient training of MFM.

W4. Additional baselines.

Table (c) compares best-of-N, flow map lookahead[1], tweedie lookahead[2], MFM, and MetaDMD on ImageNet 256×256 using SiT-XL/2. MetaDMD outperforms all other baselines on ImageReward, PickScore, and HPSv2, while performing slightly below MFM on Aesthetics. All methods are evaluated with compute budgets matched to MetaDMD’s gradient-based estimator with 16 Monte Carlo samples. For the Flow map lookahead, we use the same Decoupled MeanFlow checkpoint used to train MFM and use 16 particles with 2 clones for each particle. For the tweedie lookahead, we use 1194 euler steps for guidance. Each metric is evaluated using its corresponding reward for steering. We have added this comparison to the revised manuscript (Sec. 4.3; App. E.3; Tab. 5).

Table (c) Additional baselines. (Tab. 5, Sec. 4.3)

MethodImageReward ↑PickScore ↑HPSv2 ↑Aesthetics ↑
Teacher (no steering)0.19020.800.2404.720
Best-of-N (N=130)1.11021.400.2855.890
Flow Map Lookahead1.00021.200.2725.710
Tweedie Lookahead0.34019.090.2455.500
MFM (N=16)1.6824.600.3718.018
MetaDMD (N=16)1.71025.320.3777.920

Q1. Table 1 comparison.

Table 1 is designed as a one-to-one comparison with DMD, which many readers would be already familiar with, so that the posterior distribution matching of MetaDMD can be understood easily. The difference from MFM is illustrated in Figure 1 and mainly discussed in Sec. 2.2, especially in terms of the matching target and the resulting training cost.

Q3. Distribution versus trajectory matching.

When optimally trained, MetaDMD is guaranteed to generate samples from the posterior p1∣tmodel(⋅∣xt)p^{\text{model}}_{1|t}(\cdot \mid x_t). Its objective in Eq. (9) is the KL divergence between the student’s output distribution and this posterior, which vanishes only when the two distributions coincide (Sec. 3.1).

For our application, estimating the steering value function requires samples from the conditional endpoint distribution, which motivates MetaDMD to be trained with posterior distribution matching. Therefore matching intermediate states of the GLASS Flows ODE along the auxiliary time axis is unnecessary, and introduces additional learning requirements. Our results in Table (a) support this practical benefit by showing that MetaDMD achieves higher downstream steering performance than MFM with less training GPU hours. We therefore view distribution matching as an empirically efficient alternative for this application, rather than claiming that it is universally better than trajectory distillation.

Pathwise-consistent training and distribution-matching approaches are both important and active research directions, each with distinct strengths. DMD-style training has been successfully applied to large-scale generative models, including few-step text-to-video models [3] and interactive world models [4]. We extend this line of research by adapting DMD to train few-step posterior samplers for large-scale text-to-image models. We have added this discussion to the revised manuscript (App. C.1).

[1] Sabour et al. Test-Time Scaling of Diffusions with Flow Maps. In NeurIPS, 2026.

[2] Chung et al. Diffusion Posterior Sampling for General Noisy Inverse Problems. In ICLR, 2023.

[3] Huang et al. Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion. In NeurIPS, 2025.

[4] He et al. Matrix-Game 2.0: An Open-Source, Real-Time, and Streaming Interactive World Model. arXiv preprint arXiv:2508.13009, 2025.

Reviewer_QSd2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes MetaDMD, a framework for efficient posterior sampling in test-time steering of diffusion and flow models. Instead of distilling exact sampling trajectories as in existing stochastic flow map methods, MetaDMD extends Distribution Matching Distillation (DMD) to directly match conditional posterior distributions. This design avoids consistency losses and aims to improve training stability and efficiency. Experiments on ImageNet and Stable Diffusion 3.5 Medium demonstrate promising reward steering performance with significantly fewer training iterations compared to Meta Flow Maps (MFM).

Critical strengths
  • The paper is generally well-written and clearly motivated. The observation that posterior sampling only requires matching the final distribution rather than reproducing intermediate trajectories is reasonable and provides an interesting alternative to existing flow map distillation methods.

  • The proposed approach eliminates the need for consistency losses, Jacobian-vector products, and pretrained flow map initialization, simplifying the training procedure. The teacher-free steering variant is also a practical design that further reduces inference requirements.

  • The experiments cover both class-conditional generation and large-scale text-to-image generation, with multiple reward functions and both gradient-based and gradient-free steering. The reported improvements over MFM, particularly on SD3.5-M, are encouraging.

Critical weaknesses
  • Limited methodological novelty. While extending DMD to conditional posterior sampling is reasonable, the main distribution matching objective and training strategy largely follow existing DMD methods. The construction of the posterior denoiser in Eq. (10) also builds on the framework of GLASS Flows. Therefore, the main contribution appears to lie in combining and adapting existing techniques rather than introducing a fundamentally new distillation principle. The paper would benefit from a clearer discussion of its methodological novelty relative to DMD and stochastic flow maps.

  • Insufficient analysis of posterior sampling quality. Since the central claim is that distribution matching can replace trajectory-level distillation, it would be important to directly evaluate how accurately MetaDMD approximates the target posterior distribution. However, the experiments mainly focus on downstream reward scores, without quantitative comparisons of posterior approximation quality against GLASS Flows or MFM. Therefore, it remains unclear whether the improvements come from more accurate posterior sampling or other factors in the steering process.

  • Insufficient computational efficiency and stability analysis. The paper emphasizes that MetaDMD requires 6× fewer training iterations on SD3.5-M and 250× fewer on ImageNet compared to MFM. However, training iterations alone do not necessarily reflect actual computational costs, particularly when different training objectives and pretraining stages are involved. The paper does not report sufficient information about GPU hours, hardware configurations, or memory consumption. Moreover, the claimed training stability improvements are not directly supported by convergence curves or training dynamics comparisons.

  • Limited evaluation of reward alignment. Most experiments evaluate steering using the same reward functions employed during optimization. Improvements in these metrics do not necessarily imply better perceptual quality or human preference alignment, as reward hacking may occur. Additional cross-reward evaluations using independent reward metrics would provide stronger evidence of the effectiveness and generalization of the proposed method.

  • Inconsistent improvements under gradient-free steering. Although MetaDMD performs well on SD3.5-M, its advantages are less consistent on ImageNet. For example, Table 6 shows that MFM outperforms MetaDMD on ImageReward, HPSv2, and Aesthetic scores when using gradient-free steering with N=16. This suggests that the benefits of MetaDMD may depend on the steering estimator and sampling configuration. A more detailed discussion of these results would help clarify the strengths and limitations of the proposed approach.

Questions for authors
  1. Can the authors provide quantitative comparisons of posterior approximation quality between MetaDMD, MFM, and GLASS Flows, possibly using suitable posterior statistics or other approximation metrics?

  2. What are the actual training costs in GPU hours and peak memory usage? Could the authors clarify the hardware configurations and distinguish pretraining costs from the subsequent training costs?

  3. Can the authors provide training convergence curves or a simple comparison of training dynamics to better demonstrate the claimed stability and efficiency improvements?

  4. Could the authors provide a lightweight cross-reward evaluation to investigate whether improvements on the optimized reward also transfer to independent evaluation metrics? This would help assess the generalization of reward alignment and potential reward hacking.

I would be willing to reconsider and potentially increase my rating if the authors can adequately address the above concerns.

3 comments

We thank the reviewer for the careful reading and constructive feedback. The reviewer’s questions mainly ask whether the training of MetaDMD is truly efficient and effective. All revisions are shown in skyblue in the revised PDF, and those addressing your comments are marked with QSd2 in purple.

W1. Novelty and contributions.

Since steering only requires samples from the posterior, MetaDMD is motivated to train the posterior sampler with DMD, which matches only the endpoint distribution. To the best of our knowledge, our work is the first to train an efficient stochastic posterior sampler on a large-scale model such as SD3.5-M and to demonstrate its empirical effectiveness (Sec. 1). As shown in Table (a), MetaDMD achieves stronger steering performance than MFM on SD3.5-M, including MFM initialized from the teacher with twice the training compute and MFM initialized from a flow map with more than 2.5 times the training compute. These results support the practical value of posterior distribution matching for scaling stochastic posterior samplers to large text-to-image models.

Table (a) GenEval score on SD3.5-M, N=8, gradient-free estimator. (Fig. 3, Sec. 4.2; Tab. 11, App. F)

MethodTraining iterationsCumulative GPU-hoursGenEval score
Unsteered teacher baseline——0.672
MetaDMD00.000.688
MetaDMD1,0008.380.776
MetaDMD2,00016.750.776
MFM (Teacher initialization)00.000.556
MFM (Teacher initialization)1,00011.240.691
MFM (Teacher initialization)2,00022.470.707
MFM (Teacher initialization)3,00033.710.754
MFM (FlowMap initialization)010.220.589
MFM (FlowMap initialization)1,00013.790.661
MFM (FlowMap initialization)2,00017.370.672
MFM (FlowMap initialization)10,00045.970.714

We would appreciate it if the reviewer could consider this result as part of our contribution.

W2, Q1. Posterior accuracy.

Following the reviewer’s suggestion, we conducted a posterior recovery experiment on SD3.5-M in Table (b), and show that MetaDMD achieves lower posterior FID than MFM at all three evaluated time points.

Table (b) Posterior FID comparison (Tab. 9, App. E.1)

Methodt=0t=0.1t=0.5
MFM73.32111.83111.23
MetaDMD25.5622.4924.82

We constructed conditioning states by applying forward noising to teacher generated images at levels corresponding to t=0, t=0.1, and t=0.5. For each conditioning state, we generated 500 reference posterior samples by simulating the SDE with 200 Euler-Maruyama steps. We then computed the FID between these reference samples and 500 samples generated by MetaDMD and MFM under the same condition. Table (b) reports aggregate FID over six prompts, with one prompt selected from each of the six GenEval categories. We have added this posterior recovery experiment to the revised manuscript (Sec. 4.2; App. E.1).

W3-1, Q2-1, Q3. Efficiency and stability analysis.

MetaDMD is more computationally efficient than MFM in GPU-hours as well as in training iterations. Table (a) reports GenEval scores alongside cumulative training GPU-hours for both methods. MetaDMD achieves a GenEval score of 0.776 after 16.75 GPU-hours, whereas MFM reaches only 0.754 with teacher initialization even after twice the GPU-hours (33.71) and 0.714 with flow-map initialization even after 45.97 GPU-hours. Table (a) and Figure 3 of the revised manuscript, which plots the GenEval score at every training checkpoint, also show the training stability of MetaDMD. We have added this comparison to the revised manuscript (Sec. 4.2; App. F).

W3-2, Q2-2. Peak memory usage and hardware configurations.

Training MetaDMD for 2,000 iterations required approximately 17 GPU-hours on NVIDIA RTX A6000 GPUs. MFM and MetaDMD used 21.08 and 22.89 GiB of peak allocated memory per GPU respectively, under matched settings of SD3.5-M with four NVIDIA RTX A6000 GPUs, BF16, FSDP, gradient checkpointing, and a batch size of 2 per GPU. We have added these computational details to the revised manuscript (App. D).

W4, Q4. Cross-reward evaluation.

Table (c) and Table (d) show that MetaDMD’s improvements extend beyond the reward used for steering. Relative to the unsteered teacher, MetaDMD improves held-out reward scores in 9/12 evaluations on ImageNet and 12/12 on SD3.5-M, compared with 8/12 and 4/12 for MFM, respectively. These results suggest that MetaDMD’s gains are not confined to the optimized reward. For each of the four steering rewards ImageReward, PickScore, HPSv2, and Aesthetics, we evaluate the three remaining rewards, yielding 12 cross-reward evaluations per setting. Tables (c) and (d) report the results for ImageNet with the gradient-based estimator (N=16) and SD3.5-M with the gradient-free estimator (N=8). Bold indicates the higher held-out reward score within each MFM/MetaDMD pair. We have added these cross-reward evaluations to the revised manuscript (Sec. 4.1; App. E.2).

Table (c) ImageNet, N=16, gradient-based estimator (Tab. 10, App. E.2)

Steering rewardMethodImageReward ↑PickScore ↑HPSv2 ↑Aesthetics ↑
—Teacher baseline0.19020.800.2404.720
ImageRewardMFM1.68220.010.2565.423
MetaDMD1.71020.120.2605.350
PickScoreMFM0.62324.600.2985.378
MetaDMD0.67025.320.3055.233
HPSv2MFM0.79922.100.3715.619
MetaDMD0.94022.260.3775.550
AestheticsMFM-0.10219.170.2328.018
MetaDMD0.33019.540.2407.920

Table (d) SD3.5, N=8, gradient-free estimator (Tab. 10, App. E.2)

Steering rewardMethodImageReward ↑PickScore ↑HPSv2 ↑Aesthetics ↑
—Teacher baseline0.79022.270.2765.480
ImageRewardMFM0.82322.210.2755.456
MetaDMD1.33022.510.2955.560
PickScoreMFM0.80222.120.2735.462
MetaDMD1.08022.930.2975.580
HPSv2MFM0.82022.200.2755.482
MetaDMD1.01022.510.2975.540
AestheticsMFM0.79522.140.2745.517
MetaDMD0.94022.370.2885.830

W5. Inconsistent improvements across N.

We agree with the reviewer that MetaDMD does not uniformly outperform MFM under gradient-free steering on ImageNet. We attribute this difference to the mode-seeking tendency of DMD [1,2] (App. E.4). Image distributions conditioned on detailed text prompts are likely to have fewer modes than those conditioned on class labels, which cover a much broader range of images. The mode-seeking behavior of DMD is therefore less likely to affect steering on text-to-image models such as SD3.5-M, as supported by the low posterior FID in Table (b) and the steering results on this model, whereas it can be detrimental for class-conditional generation on ImageNet. We believe a large text-to-image model is closer to practical real-world use, and we would appreciate the reviewer’s consideration of this point.

[1] Lu et al. Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis. In ICCV, 2025.

[2] Yin et al. One-step Diffusion with Distribution Matching Distillation. In CVPR, 2024.

Dear Reviewer QSd2,

Thank you again for your careful review and constructive feedback. As the author–reviewer discussion period ends on November 18 (AoE), we would be grateful if you could let us know whether our responses have addressed your concerns.

Following your suggestions, we compared the training cost with MFM in GPU-hours, including MFM initialized from the teacher, and reported the hardware and peak memory usage. We also added posterior FID to directly measure posterior accuracy, cross-reward evaluations with held-out rewards, and a discussion of the gradient-free results on ImageNet. All of these are incorporated into the revised paper, and we would be happy to clarify any remaining questions.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of revisions

We thank all reviewers for their constructive feedback. All revisions are shown in skyblue in the revised PDF and tagged with the reviewer ID (QSd2, qA8B, cdL2).

1. Contributions [QSd2]

  • First stochastic posterior sampler trained efficiently on a large T2I model, outperforming MFM even with more training compute: Sec. 1

2. Clarification of the method

  • Why matching at multiple noise levels is not a trajectory constraint [cdL2]: Sec. 3.1
  • Distribution versus trajectory matching, with correctness at the optimum [qA8B]: Sec. 3.1, App. C.1
  • Relation to DMD2 and omission of the adversarial loss [cdL2]: App. D

3. New experiments

  • Training cost compared in GPU-hours instead of training iterations throughout the paper [QSd2, qA8B]: Abstract, Sec. 1, Sec. 4.1, Sec. 4.2, Sec. 4.3, Sec. 6
  • Training efficiency in GPU-hours on SD3.5-M, including MFM initialized from the teacher [QSd2, qA8B, cdL2]: Sec. 4.2, Fig. 3, App. F, Tab. 11
  • Training cost in GPU-hours on ImageNet [QSd2, qA8B]: Sec. 4.3
  • Hardware and peak memory usage [QSd2]: App. D
  • Posterior accuracy measured by posterior FID [QSd2]: Sec. 4.2, App. E.1, Tab. 9
  • Cross-reward evaluation [QSd2]: Sec. 4.1, App. E.2, Tab. 10
  • Gradient-based steering with DPS on SD3.5-M [qA8B]: Sec. 4.2, App. E.4, Tab. 4
  • Comparison with Weighted Diamond Maps [qA8B]: Sec. 4.2, App. E.5
  • Comparison with best-of-N, flow map look-ahead and Tweedie look-ahead [qA8B]: Sec. 4.3, App. E.3, Tab. 5

4. Additional discussion

  • Gradient-based versus gradient-free steering on SD3.5-M [qA8B]: Sec. 4.2, App. E.4
  • Gradient-free steering on ImageNet [QSd2]: App. E.4

5. Clarification of the experimental setup

  • MFM as the representative of stochastic flow maps, including Posterior Diamond Maps [qA8B]: Sec. 4.1
  • ImageNet evaluation protocol with the official MFM-XL/2 checkpoint [qA8B]: Sec. 4.1, App. D
  • Reference to the gradient-free ImageNet results alongside Tab. 3 [qA8B]: Sec. 4.3, Tab. 8
AuthorsOpenReview ↗

Thank you to the AC and reviewers

We sincerely thank the Area Chair for managing the review of our submission and all reviewers for their thoughtful feedback throughout the discussion. The reviewers' suggestions led to new experiments and clarifications that have significantly improved the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.