acceptodds
Under review as a conference paper at ICLR 2027

Unifying Distributional Training for One-Step Visual Generation

11 views

Abstract

Distributional training provides collective supervision for one-step visual generation by matching real and generated features in frozen representation spaces. We introduce a unified theoretical framework that separates distribution modeling from matching discrepancy and connects global objectives to pointwise feature updates through Wasserstein gradient flow. Under this framework, FD-Loss and Gaussian-kernel Drifting are recovered through Gaussian optimal transport and kernel-density-based KL matching respectively. The framework motivates MGFlow, which models feature distributions with Gaussian mixtures at an adjustable granularity between global moments and sample-based representations. MGFlow supports both optimal transport and score based matching, and couples mass-constrained sample assignment with paired component updates to address mode collapse that mixture expressivity alone does not resolve. On ImageNet , MGFlow substantially surpasses the FD-Loss baseline, achieving state-of-the-art results with 1.45 on pMF-H and 1.64 on JiT-H. For text-to-image generation, MGFlow post-trains FLUX.2 [klein] 4B into a one-step generator that outperforms the original four-step model on both GenEval and PickScore. Project page: https://mgflow-repo.github.io.

Primary area:
generative models
Keywords:
Generative models, Image Generation, Diffusion Models, One-step Generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_yN9UOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents a unified framework for distributional training of one-step visual generators, separating distribution modeling from the discrepancy used for matching. Within this framework, the authors propose MGFlow, which models feature distributions using Gaussian mixtures and combines mass-constrained sample allocation with paired component updates. Both optimal-transport-based and KL-based variants are developed. Experiments on ImageNet across multiple generator backbones and text-to-image generation demonstrate the effectiveness of the approach, with additional evaluations using held-out feature encoders.

Critical strengths
  1. A useful conceptual framework supported by a coherent algorithmic design. Separating distribution modeling from matching discrepancy clarifies the relationships between existing approaches and provides a systematic way to develop alternatives. MGFlow makes this framework concrete: mass-constrained allocation enforces reference component proportions in the assigned samples, while paired updates preserve the correspondence between allocation and transport. The ablations support the importance of combining these two mechanisms.

  2. Substantial empirical validation beyond the conceptual formulation. The ImageNet experiments cover multiple backbones and model sizes, and improvements on held-out feature encoders provide evidence that the benefits extend beyond the representations used for training. The appendix also includes detailed implementation, ablation, and computational analyses. Together, these results strengthen the case that the proposed framework leads to a useful training method.

Critical weaknesses
  1. The image-level evidence for improved mode coverage remains indirect. The toy experiment and assignment ablations support the proposed mechanism, but aggregate distributional scores do not distinguish better mode proportions from better within-mode fidelity. Since correcting mass imbalance is central to the motivation, a direct coverage diagnostic on generated images would strengthen the empirical justification.

  2. Increasing the number of mixture components is not straightforward. Appendix I.2 shows that some components receive too few samples to reliably estimate their full covariances, while longer EMA windows slow adaptation to the changing generator. This limits the granularity of the mixture representation and leaves its scalability to distributions requiring many more components uncertain.

Questions for authors
  1. Clarification of the theoretical scope: Appendix B.5 appropriately distinguishes continuous-flow energy descent from a finite-step guarantee for generator optimization. For practical MGFlow-KL, does any descent interpretation also survive the combination of LP reassignment, EMA statistics, and covariance regularization? If not, could the authors make this boundary more explicit in the main text?

  2. Evidence supporting the proposed mechanism: Can the authors provide a direct diagnostic on generated images comparing mode proportions or coverage under posterior assignment, LP allocation with global updates, and LP allocation with paired updates? This would help determine whether the improvements specifically reflect better mass allocation, rather than primarily better within-mode fidelity.

3 comments
AuthorsOpenReview ↗

Response to Reviewer yN9U

Thank you for recognizing the value of our unified framework, the coupled allocation-and-update design, and the breadth of the empirical validation. We address the requested coverage diagnostic and clarify the statistical and theoretical scope below.

[W1, Q2] A direct diagnostic of generated images under different assignments

The assignment ablation in Table 2 reports generation quality across the three variants. As a generation metric, Fréchet Distance is itself sensitive to mode coverage, not only in-mode fidelity, and thus lower FD already reflect better mass allocation. To further test mass allocation directly, we add a complementary component-share diagnostic with JiT-B trained on SIM encoder spaces. LP-paired most closely matches the full 16-component share vector: it has the lowest TV and the smallest absolute share error on 15 of the 16 components at the reported precision.

We compare Posterior, LP-global, and LP-paired using K=16, the same initialization, and 10 training epochs. Our reference Inception Gaussian mixture evaluates 50,000 newly generated one-step images and 50,000 real images. Each component's soft share is the average posterior probability. Crucially, evaluation imposes no LP quotas: it measures where generated images actually fall under the reference GMM.

We sort the components once by decreasing real-data share and show ranks 1, 4, 8, 12, and 16, spanning the largest to the smallest components. Every method is evaluated against the same component identities. Values are percentages.

Ranked ComponentRealLP-pairedLP-globalPosterior
110.6411.211.540.87
49.178.960.284.69
88.097.860.0211.20
122.882.8897.514.00
160.140.000.000.20

Across all 16 components, the Total Variation Distance (TV) is 0.21682 for Posterior, 0.94625 for LP-global, and 0.03782 for LP-paired. The maximum absolute component-share deviation is approximately 1.28 percentage points for LP-paired, compared with 9.77 for Posterior and 94.63 for LP-global. This shows why training-time quotas alone are insufficient and why preserving the allocation in paired update matters.

[W2] Statistical support when increasing the mixture component count

We thank the reviewer for raising this point. We would like to clarify that this limitation primarily stems from the scale of ImageNet, where increasing the number of mixture components reduces the samples available for reliable covariance estimation, rather than from MGFlow itself. In fact, this valuable comment highlights an exciting direction for future research on finer-grained mixture representations using substantially larger, industry-scale datasets. We will discuss this promising direction in the appendix.

[Q1] Descent under LP reassignment, EMA, and covariance regularization

A local descent interpretation survives LP assignment and covariance regularization; EMA makes the current-moment identity approximate while providing a more stable population estimate. For each global batch and encoder/branch, Algorithm 3 solves the LP once and keeps the assignment fixed throughout that update. We then fix the assignment matrix RR and define the paired KL using the current R-weighted means and covariances of the generated features ZZ, with λI\lambda I added to both reference and generated covariances:

FRλ=∑kπkKL(N(μk(Z),Σk(Z)+λI)∥N(μp,k,Σp,k+λI))F^\lambda_R = \sum_k \pi_k \text{KL}(\mathcal{N}(\mu_k(Z), \Sigma_k(Z) + \lambda I)\Vert\mathcal{N}(\mu_{p,k}, \Sigma_{p,k} + \lambda I))

Direct differentiation gives:

∇zFRλ=−vpair,n/B\nabla_{z} F^\lambda_R = -v_{\text{pair},n}/B

This identity holds for any fixed assignment satisfying the mass constraints, including LP assignments; it does not require differentiating through the assignment solver. Covariance regularization changes the objective to the KL between regularized Gaussians while preserving this gradient relationship. Combining this identity with Eq. (46), the detached-target regression gradient equals η times the objective’s parameter gradient at the current iterate. A sufficiently small gradient-descent step therefore decreases this fixed-assignment objective under standard smoothness conditions.

EMA introduces a deliberate trade-off between tracking the current distribution and reducing minibatch noise in its moment estimates. By aggregating statistics across batches, it provides a larger effective sample population without requiring a correspondingly large gradient batch. This strategy has empirical support in FD-Loss Sec 4.1, where the EMA estimator performs well across a broad range of decay rates and its best setting outperforms the tested queue-based alternatives.

We will add the explicit claims to Appendix B.5 regarding assignment algorithms and covariance regularization, and clarify the theoretical boundary of EMA updates in the main text.

AuthorsOpenReview ↗

Follow-up on our rebuttal

Dear Reviewer yN9U,

Thank you again for your thoughtful comments. Your questions prompted us to measure component coverage directly and make the descent argument more explicit. The rebuttal reports component shares across all 16 components, discusses the sample requirements as the number of components grows, and explains the local descent interpretation under fixed LP assignments.

If the response addresses your concerns, we would be grateful if you would consider raising your score. We would be happy to elaborate on any point that would benefit from further discussion.

Thank you for the detailed rebuttal and follow-up clarification. I appreciate the authors' efforts to address my concerns. Overall, the explanation has addressed my most questions regarding component coverage and the theoretical scope of the descent argument. Considering the paper’s overall insights and novelty, I remain inclined to maintain my original rating of weak accept

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

This paper presents a unified view of distributional training and proposes MGFlow for one-step visual generation. MGFlow models feature distributions with Gaussian mixtures and combines sample allocation with paired Wasserstein or score-based updates. Experiments cover ImageNet post-training and text-to-image generation. MGFlow-KL achieves FDr₆ of 1.45 on pMF-H and 1.64 on JiT-H, and a GenEval score of 0.900 after post-training FLUX.2 [klein] 4B.

Critical strengths
  1. Figure 3 and Table 2 in the paper show the benefits of paired updates, where LP-paired greatly improves over LP-global. This is a concrete innovation aside from simply replacing a single Gaussian with a Gaussian mixtures.

  2. The improvements in the paper involve multiple architecture sizes and representation spaces. For example, on pMF-H, MGFlow-KL improves FDr₃ from 2.69 to 1.88 over FD-Loss. This shows its benefits are not specific to one architecture or representation.

Critical weaknesses
  1. The main improvements are not sufficiently shown in ablations. Table 6 shows that only changing the training recipe can improve pMF-B FDr₆ from 3.50 to 3.17, compared with 3.09 for MGFlow-KL. However, the equivalent full-training controls are missing for larger architectures. Similarly, the ablations that compare with methods such as W-Flow are restricted to 10-epoch pMF-B training with a single Inception encoder, but the main results use 100 epochs and multiple encoders. Controlling the epochs alone is not exactly budget matching, and what is more interesting would be running til convergence or running with a larger number of epochs. Comparisons that control the actual computational costs would strengthen the main claims.

  2. The paired score updates may cause additional bias, since the reference components are fitted by EM, but generated components use LP. This mismatch may lead to different component moments even for identical real and generated distributions, as acknowledged in the paper. Therefore, the updates may move samples away from an already matched distribution. The paired-KL upper bound does not address this inconsistency.

Questions for authors

Could the authors address the concerns raised in the Weaknesses section, particularly the need for aligned baselines and fuller comparisons, and the potential target inconsistency caused by the EM/LP assignment mismatch?

4 comments
AuthorsOpenReview ↗

Response to Reviewer uDtw, Part 1

Thank you for recognizing paired updates as a concrete algorithmic contribution and the improvements across architectures and representation spaces. We have extended the aligned controls, added a 100-epoch SIM comparison and an equal-compute evaluation, and examined the EM/LP consistency concern directly.

[W1, Q1] Ablation studies in broader settings.

We extend the single-Gaussian FD (Aligned FD-Loss) control to JiT-B, pMF-L, and JiT-L, matching the pretrained backbone, three frozen SIM encoders, fixed reference normalization, and statistics-update recipe. All runs use 100 epochs; lower FDr⁶ is better.

BackboneAligned FD-LossMGFlow-KL
pMF-B3.173.09
JiT-B4.553.70
pMF-L1.851.74
JiT-L2.881.92

The improvement remains on all four backbones after recipe alignment, including both larger architectures. The gain is larger on JiT, where the distributional structure is most valuable. FD-Loss directly optimizes the FD evaluation metric, while MGFlow does not; the per-encoder breakdown in our response to Reviewer zhGo [W1] shows that the additional components improve distribution matching across held-out perspectives rather than optimizing only the first two global moments.

We also rerun the Table 1 model-discrepancy ablation for 100 epochs with SIM encoders:

OT-based MethodResultScore-based MethodResult
GaussianFD-Loss3.17Gaussian KL3.14
GMMGFlow-W23.03MGFlow-KL3.09
Sample-basedW-Flow3.79Drifting3.90

The Table 1 ablation holds after extending training to 100 Epochs and using multiple encoders. Within each discrepancy family, MGFlow achieves the best performance as Gaussian Mixtures models the feature distribution with continuity and flexible granularity.

[W1, Q1] Whether the gains remain under equal computational cost.

We evaluate checkpoints under a common budget of approximately 156.16 H100 GPU-hours (100 epoch for the fastest method). All variants use pMF-B:

MethodEqual-compute epochsH100 GPU-hoursEqual-compute FDr⁶ ↓Equal-compute FID ↓
Aligned FD-Loss75.48156.143.351.86
Gaussian KL, K=1100.00156.163.141.69
MGFlow-W₂, K=1+460.44156.103.201.84
MGFlow-KL, K=1+4+1656.32156.153.241.67
W-Flow63.16156.153.862.35
Gaussian-kernel Drifting80.32156.143.952.81

At matched compute, Gaussian KL achieves the best FDr⁶ and FID among the compared methods—and is itself a training method introduced in this paper, derived directly from our distributional training framework. This result demonstrates the framework’s value in identifying an efficient alternative to Gaussian FD. Mixture modeling provides a complementary benefit: MGFlow improves both FDr⁶ and FID over all existing baselines in this comparison, while using substantially fewer training epochs. The framework therefore delivers efficient formulations under a fixed compute budget.

Moreover, the relative cost of mixture modeling also decreases as the generator grows. At fixed feature dimension, component count, and batch size, the additional statistics, LP, and score computations are largely independent of the generator’s parameter count, while neural computation becomes more expensive. For FLUX.2 [klein] 4B, adding the K=4 branch increases the measured statistics/LP/score/feature-communication stage from 0.37 to 1.10 seconds: an additional 0.73 seconds, equivalent to approximately 1.6% of a 45-second training step, compared to the 66% on pMF-B. This makes finer distributional modeling increasingly affordable on larger generators. These results establish a practical scaling advantage for MGFlow, which adds richer distributional supervision at a diminishing relative cost as the generator scales.

AuthorsOpenReview ↗

Response to Reviewer uDtw, Part 2

[W2, Q1] Whether EM fitting and LP allocation preserve a matched target distribution.

We thank the reviewer for raising this question. The key distinction is between exact invariance of a single-batch estimator and the component alignment reached by the complete training procedure. The reference GMM is fitted to numerical convergence with full-data EM. During training, generated-side statistics are estimated from finite batches, updated after each generator step, and accumulated with EMA.

The reviewer’s question can therefore be tested at three levels: a full-data EM update, a single-batch EM/LP update, and the generated-side statistics saved by the trained model. We use the fitted Inception K=16K=16 reference and the complete feature bank. The same 1,024 real features are used for the batch EM and LP comparisons:

Statistics estimateNormalized mean shiftRelative covariance shiftWeight L1 shift
Full-data EM0.002240.015380.00291
Batch 1024 EM0.136951.300320.09504
Batch 1024 LP0.129491.27078≈0
Trained pMF-B checkpoint0.01120.0425≈0

The full-data EM update is nearly stationary, confirming the consistency of the reference fit. With either EM or LP assignment rules, a single finite batch produces large moment variation because 1,024 samples provide limited support for 2,048-dimensional full covariances. LP preserves the prescribed component masses exactly and even gives slightly smaller raw mean and covariance shifts on this batch, proving that LP is a more stable allocator under finite-batch estimation.

The trained checkpoint gives the direct practical answer to the reviewer’s concern. Its generated-side component means and covariances are much closer to the reference than either raw batch estimate: the mean and covariance discrepancies are approximately 11.6× and 29.9× smaller than the batchwise LP shifts. Thus, the trained statistics do not remain at the large moment discrepancy produced by a single LP batch. This improvement comes from the complete training dynamics. Paired updates move generated features toward their assigned reference components, while EMA aggregates statistics across many successive batches instead of replacing the estimate with one noisy batch. The resulting generated-side GMM therefore benefits from a substantially larger effective observation history.

AuthorsOpenReview ↗

Follow-up on our rebuttal

Dear Reviewer uDtw,

Thank you again for your careful review. Your suggestions helped make the comparisons more informative. We added aligned FD-Loss baselines on the requested backbones, extended the SIM comparison to 100 epochs and matched compute, and examined the EM/LP moment mismatch directly.

If these additions address your concerns, we would greatly appreciate your considering a higher score. If any questions remain, we would be glad to discuss them further. Thank you for your time.

Thank you for the response and the additional results. Most of my concerns have been addressed. I decided to maintain my positive score.

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

This paper focuses on one-step visual generation using distributional training, which is a training objective that matches distributions of real and frozen representation spaces rather than only sample-level. The paper provides a unified theoretical framework for distributional modeling, which shows that existing methods such as FD-Loss and Gaussian-kernel Drifting can be recovered from it. The work also proposes a method motivated by the framework called MGFlow, which models distributions using Gaussian mixtures instead of a single Gaussian or individual sample-centered kernels, and supports both OT and score-based matching. On two architectures on ImageNet, the work shows that MGFlow achieves SOTA results, and can also be used to post-train Flux.2 into a one-step generator with strong performance.

Critical strengths
  • The paper proposes a clean unified framework for distributional training that distinctly separates how feature distributions are modeled from how their discrepancy is measured. Using Wasserstein gradient flow, the authors establish a common formulation connecting FD-Loss to Gaussian modeling with W2 matching, and Gaussian-kernel Drifting to kernel-density modeling with KL matching. This provides a useful foundation for understanding existing methods and for designing future methods.
  • The framework naturally motivates MGFlow, which uses Gaussian mixtures to capture finer distributional structure. Sections 3 and 4.1 support the design decisions through ablations controlled experiments by taking a look at component assignment, and combinations of hierarchical coarse and fine mixture models. The results suggest that the method performs strongly compared to existing baselines.
Critical weaknesses
  • The quality improvements in Table 4 are metric-dependent. Both MGFlow variants have worse FID than FD-Loss across all six ImageNet configurations, and IS also decreases on the JiT models. I understand that FDr6 can provide a broader evaluation especially with respect to distribution matching than FID alone, but the paper lacks analysis of what is exactly gained and lost in this tradeoff. Table 19 also shows that improvements across held-out encoders are uneven: ConvNeXt performance worsens on the pMF models, while the pMF-H improvement in held-out FDr3 is largely driven by CLIP.
  • The main quality results do not report error bars, confidence intervals, or variation across independent training runs. This makes it difficult to assess the reliability of smaller improvements over the strongest baselines, particularly the PickScore improvement from 21.85 to 21.98. The absence of uncertainty estimates limits confidence in these close rankings.
Questions for authors

Clarification requests:

  • Have the main results been evaluated across independent training runs or multiple generation seeds? In particular, are there confidence intervals for the PickScore differences on the 499 evaluation prompts? Given the relatively small improvement from 21.85 to 21.98, this would help determine how much weight to place on the reported advantage.

Additional Experiments:

  • Tables 1 and 3 provide useful single-Gaussian versus mixture comparisons in the 10-epoch, Inception-only setting. Could you provide similar ablations for same 100-epoch, three-encoder setting used for the main results, particularly a (K=1) Gaussian-KL baseline against MGFlow? This would help us ablate the multi-Gaussian component more strongly
  • Although the GenEval reference-data construction follows prior work, the post-training reference set includes images generated from the same GenEval prompts used for evaluation, which limits generalization. Do the improvements also hold without GenEval-derived references or on a separate set of compositional prompts or additional benchmarks (aside from GenEval/ PickScore with very small variation in performance across models)? This would help establish generalized performance gains.
4 comments
AuthorsOpenReview ↗

Response to Reviewer zhGo, Part 1

Thank you for recognizing the clarity of our unified framework and the empirical support for MGFlow’s design. Your comments help us sharpen the interpretation of the results and strengthen the experimental controls.

[W1] Interpreting the FDr gains alongside FID and IS

We acknowledge the FID and IS trade-off in Table 4 and provide a per-encoder breakdown to explain the improvement in FDr. We compare FD-Loss with fixed FD(R,V) normalization against MGFlow-KL on JiT-B and pMF-B. Both train on SIM for 100 Epochs. Entries show normalized FD (↓) from FD-Loss to MGFlow-KL (FD-Loss → MGFlow).

Encoder or scoreJiT-BpMF-B
SIM · Inception0.96 → 0.761.08 → 0.93
SIM · MAE1.27 → 1.751.72 → 1.99
SIM · SigLIP1.89 → 2.603.28 → 4.35
Held-out · ConvNeXt2.11 → 1.221.35 → 1.09
Held-out · DINOv25.50 → 4.623.73 → 3.59
Held-out · CLIP15.59 → 11.267.83 → 6.58
FDr³7.73 → 5.704.30 → 3.76
FDr⁶4.55 → 3.703.17 → 3.09

Our aligned FD-Loss with fixed FD(R,V) normalization achieves FDr⁶ of 4.55 on JiT-B and 3.17 on pMF-B. The published SIM FD-Loss with FD(R,G) normalization reports 5.53 and 3.50, respectively. This gives a stronger single-Gaussian baseline than reported FD-Loss for the MGFlow-KL comparison. (Yang et.al., Table 3)

FD-Loss directly optimizes global FD in SIM and thus has a lower mean FDr in SIM encoder spaces on both backbones (1.37→1.70 and 2.03→2.42). MGFlow-KL improves all three held-out distances on both backbones, driving the reductions in FDr³ and FDr⁶, demonstrating better overall performance and generalization to non-training encoder spaces. FID improves (1.61→1.28 and 1.82→1.56), while IS falls (333.3→308.5 and 339.4→323.0). Published FD-Loss still leads on FID and IS in Table 4. The held-out gains support transfer across feature spaces, proving that additional components help fit distribution across all perspectives, rather than overfitting the first and second momentums.

[W2, Q1] Uncertainty and stability across seeds

The PickScore advantage is consistent across all five paired generation seeds, and the paired 95% confidence interval excludes zero. We regenerated all 499 prompts with five paired seeds. One-step MGFlow scores 21.973 ± 0.017, compared with 21.903 ± 0.012 for four-step FLUX.2 [klein] (mean ± seed SD). The mean paired gain is +0.070, with a 95% prompt-and-seed bootstrap interval of [0.019, 0.120]. All five seed differences favor MGFlow. This repeated evaluation supports the reliability of the positive PickScore gain. We will report this paired re-evaluation and state the scope of the uncertainty estimate explicitly.

AuthorsOpenReview ↗

Response to Reviewer zhGo, Part 2

[Q2] Checking the Table 1 and Table 3 ablations in broader settings

We provide additional ablations by extending the six-method pMF-B comparison in Table 1 from 10 to 100 Epochs with SIM (SigLIP, Inception and MAE) multiple encoder spaces. Both columns report final one-step FDr⁶ (↓).

OT-based MethodResultScore-based MethodResult
GaussianFD-Loss3.17Gaussian KL3.14
GMMGFlow-W23.03MGFlow-KL3.09
Sample-basedW-Flow3.79Drifting3.90

The Table 1 ablation holds after extending training to 100 Epochs and using multiple encoders. Within each discrepancy family, MGFlow achieves the best performance as Gaussian Mixtures models the feature distribution with continuity and flexible granularity.

For Table 3, The above table already proves the gain of increasing Gaussian components (3.17 → 3.03, 3.14 → 3.09). We additionally train JiT-B and JiT-L with SIM encoder spaces for 100 epochs, for evidence in more architectures and model sizes. Both columns report FDr⁶ (↓).

SIM backboneGaussian KL (K=1)MGFlow-KL (K=1+4+16)
JiT-B3.863.70
JiT-L2.121.92

Adding the K=4 and K=16 branches lowers FDr⁶ on both backbones: from 3.86 to 3.70 on JiT-B and from 2.12 to 1.92 on JiT-L. This extends the Table 3 component-count finding to broader settings, confirming the effect of multi-Gaussian component for distribution modeling across different architectures and in longer training steps.

[Q3] Generalization beyond GenEval-derived references

We acknowledge that the GenEval reference data limits generalization for T2I evaluation and therefore provide further evaluations beyond GenEval. We compare the one-step MGFlow joint checkpoint with the original four-step FLUX.2 [klein] checkpoint on 2 other benchmarks:

DPG-Bench: we evaluate MGFlow and FLUX.2 [klein] for 5 individual repeats. In each repeat, we generate four 512 × 512 images for each of the 1,065 prompts and score them using mPLUG-Large under the DPG-Bench protocol. The prompt set is completely different from GenEval. MGFlow scores higher than 4-step FLUX.2 [klein] in all five repeats, with a mean paired gain of +0.806 points (95% CI [0.296, 1.310], jointly resampling repeats and prompts).

ImageBench: we evaluate both models on 192 ImageBench prompts with no exact matches to GenEval prompts, generating one 512 × 512 image per prompt per model. The one-step model has comparable overall performance and higher visual preference.

CheckpointDPG-BenchImageBench Overall ↑ImageBench EPS ↑ImageBench Capability ↑
FLUX.2 [klein], four steps85.7354.1959.9548.44
MGFlow, one step86.5454.4663.0945.83

Together, these results demonstrate superior performance of 1-step MGFlow generation over the FLUX.2 [klein] 4-step generation on generalized scenarios.

AuthorsOpenReview ↗

Follow-up on our rebuttal

Dear Reviewer zhGo,

Thank you again for your thoughtful review. Your questions about the metric trade-offs and generalization helped us strengthen the evaluation. In the rebuttal, we break down the results by encoder, report uncertainty across five paired PickScore runs, and add DPG-Bench and ImageBench evaluations.

If the new results address your concerns, we would be grateful if you would consider raising your score. We would be happy to clarify any points that remain open. Thank you for taking the time to consider our response.

AuthorsOpenReview ↗

A brief follow-up on our rebuttal

Dear Reviewer zhGo,

If you have a chance to look at our rebuttal, we would be grateful for any further feedback, particularly on whether the encoder-level results and the evaluations beyond GenEval address your questions. We would be happy to clarify any points that remain open.

Thank you again for your time and thoughtful review.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of reviews and author response

Dear Area Chair,

Thank you for overseeing the discussion. We are grateful for the three reviewers’ careful feedback. Across the reviews, they highlighted the clarity of the Wasserstein gradient flow framework connecting FD-Loss and Gaussian-kernel Drifting, the contribution of coupling LP allocation with paired updates, and the empirical gains across backbones and held-out representation spaces.

We addressed the concerns raised in the reviews with the following results and clarifications.

  • Metric interpretation and generalization (zhGo). The encoder-level results show that MGFlow-KL improves all three held-out distances over aligned FD-Loss on both JiT-B and pMF-B. We also added the requested Gaussian KL control, reported uncertainty across five paired PickScore runs, and evaluated on DPG-Bench and ImageBench beyond the GenEval prompts. On DPG-Bench, the one-step model improves over the original four-step model by 0.806 points, with a 95% paired bootstrap interval of [0.296, 1.310].
  • Aligned baselines and compute (uDtw). We added aligned FD-Loss on JiT-B, pMF-L, and JiT-L, extended the six-method SIM comparison to 100 epochs, and reported matched-compute comparisons. On JiT-L, MGFlow-KL improves FDr⁶ from 2.88 to 1.92 against the aligned baseline. We also examined the EM/LP moment mismatch using full-data, single-batch, and trained-model statistics.
  • Component coverage and descent (yN9U). We measured coverage across all 16 components on JiT-B with SIM. LP-paired reduces the total variation from 0.21682 with posterior assignment and 0.94625 with LP-global to 0.03782. We also discussed the sample requirements for fitting more components and gave a local descent result for regularized paired KL under fixed LP assignments, with a separate explanation of how EMA affects that interpretation.

We believe these additions address the concerns raised in all three reviews. We would appreciate your taking the new evidence into account in your assessment. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.