acceptodds
Under review as a conference paper at ICLR 2027

Are We Overthinking One-Step Generation?

18 views

Abstract

Diffusion and flow matching are now dominant frameworks for generative modeling in continuous domains, but they typically require multiple network evaluations at inference. This has motivated extensive research on one-step generative modeling primarily built on diffusion and flow, though existing approaches often rely on complex training machinery. In this work, we introduce **Priming Models**, a simple two-stage framework for training one-step generators. Motivated by the difficulty of distributional training from scratch, Priming Models are first trained on a reconstruction objective, e.g., pixel-space denoising at a *fixed* noise level, before undergoing feature-space distributional post-training. Our central scientific finding is that this simple reconstruction stage is already enough to entirely alleviate the difficulty of distributional training. Empirically, across class-conditional image generation, text-to-image generation, and world modeling, Priming Models significantly advance the Pareto frontier of training compute vs. generation quality. On ImageNet at resolution, we achieve a state-of-the-art FID of among one-step generators using approximately one-tenth the training compute of prior methods with comparable FID.

Primary area:
generative models
Keywords:
generative models, diffusion models, flow matching, few-step modeling, one-step modeling

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

This paper studies a two-stage approach for one-step generation: reconstruction pretraining followed by distribution-matching fine-tuning. The main finding is that pretraining at a single noise level can already provide a good initialization. The paper also tests different combinations of pretraining and fine-tuning objectives and finds a recipe that achieves strong results. I think the paper has useful findings and a strong training recipe. My main concerns are the overclaims, the lack of a controlled comparison for the noise distribution, and the unclear FLOPs accounting. I would be willing to raise my score if the authors address these points.

Critical strengths
  1. The paper shows that single-noise pretraining works for one-step generation. I think this is an interesting finding, since it suggests that we do not need to learn the full denoising process to get a useful initialization.

  2. The ablations show that different pretraining objectives and distribution-matching losses can work well together. They provide useful evidence that this two-stage approach is efficient and can achieve reliable results.

  3. The authors find a training recipe that achieves state-of-the-art results on the reported benchmarks. The results benefit substantially from perceptual supervision, as shown by the ablation. However, many existing works also use such losses, so I do not consider this a separate weakness.

Critical weaknesses
  1. I think the paper overclaims its novelty. The method relies on a combination of many existing tools, including pretraining, feature-space distribution matching, multiscale features, OT-based objectives, and perceptual losses. The paper does show that these components work well together, and this is valuable. But this does not automatically make the combination an entirely new method. The “Priming Models” framing should be more precise about what is actually new.

  2. Fixed-noise denoising pretraining followed by downstream fine-tuning also has prior work, including Brempong et al. [1] for semantic segmentation and Zaidi et al. [2] for molecular property prediction. These papers should be cited and discussed. They are not about one-step generation, but they do show that fixed-noise pretraining itself is not new.

  3. I also think “entirely alleviate the difficulty of distributional training” is too strong. The experiments show that pretraining helps in the tested settings. They do not show that the difficulty has been entirely removed. I see the main contribution as showing that single-noise pretraining provides a good initialization for one-step generation, backed by useful ablations and a strong training recipe.

  4. Single-noise pretraining works, but why prefer it? The paper does not clearly answer this question. Multiple-noise pretraining performs better in the text-to-image experiments, although it also uses a different training budget. The supplementary ImageNet comparison uses different settings from the main experiment. I would like to see a comparison under the main ImageNet settings, with only the noise distribution changed. This would tell us whether single-noise pretraining has an advantage or is simply one choice that works.

  5. I am not sure what is included in the reported EFLOPs. Does the calculation include online generation of both active and reference samples? Does it include feature extraction for real and generated samples, and backpropagation through the frozen feature encoders to the generated images? These costs matter for the efficiency comparison and should be counted consistently across methods.

Questions for authors
  1. Could the authors repeat the main ImageNet experiment and only replace the fixed noise level with reweighted logit-normal? Please keep the architecture, training budget, perceptual losses, fine-tuning settings, and evaluation the same.

  2. Could the authors explain the EFLOPs calculation, including the costs mentioned in W3, and clarify whether the baselines use the same accounting?

9 comments

We thank the reviewer for the valuable feedback and suggestions. We address the concerns below.

[W1 + W3] Novelty and contribution

I think the paper overclaims its novelty. The method relies on a combination of many existing tools, including pretraining, feature-space distribution matching, multiscale features, OT-based objectives, and perceptual losses

I also think “entirely alleviate the difficulty of distributional training” is too strong.

We would like to clarify that our main contribution is to demonstrate that combining simple, established techniques enables effective and efficient one-step generation. We do not claim novelty for the individual components in our manuscript.

We believe this finding constitutes a substantial contribution in its own right. In particular, it is striking that a combination of long-established tools (such as denoising pretraining and Sinkhorn-based distribution matching) can achieve such strong empirical performance. The novelty lies in identifying a simple training recipe and demonstrating its effectiveness.

Regarding the phrase “entirely alleviating the difficulty of distributional training,” we agree that this wording is too strong. We will revise it to “substantially improving the efficiency of distributional training” to better reflect the scope.

[W2] Prior work on denoising pretraining

Fixed-noise denoising pretraining followed by downstream fine-tuning also has prior work, including Brempong et al. [1] for semantic segmentation and Zaidi et al. [2] for molecular property prediction.

We would like to point out that denoising pretraining itself has been widely used and has become a conventional technique in representation learning, starting with denoising autoencoders (Vincent et al., 2008).

The works by Brempong et al. and Zaidi et al. also demonstrate its usefulness for downstream discriminative tasks, including semantic segmentation and molecular property prediction.

Crucially, our work uses denoising pretraining for generative modeling, whose purpose significantly differs from conventional representation learning tasks.

[W4 + Q1] noise distribution

Following the reviewer’s suggestion, we conducted an additional experiment under the Priming-B-80 setting in ImageNet, replacing the fixed noise level with a JiT-style reweighted logit-normal distribution while keeping everything else unchanged.

PretrainingFID ↓FDr⁵ ↓
Fixed 0.81.4712.71
JiT-style reweighted logit-normal1.5213.81

Under this controlled setting, single noise level pretraining at s=0.8 performs better than reweighted logit-normal pretraining on both FID and FDr^5. This result supports the effectiveness of single noise level pretraining, while also showing that a noise distribution that doesn’t cover the entire noise level is not necessarily beneficial for our training recipe. We will include this comparison in the revised paper.

We further provide an intuition for why a single, carefully selected noise level can be effective. Prior works have highlighted the importance of noise-level selection in diffusion model training (Choi et al., 2022; Hang et al., 2023, 2025). In particular, Hang et al. (2025) show that emphasizing noise levels around log SNR = 0 can improve training efficiency. Furthermore, Hoogeboom et al. (2023) and Chen (2023) demonstrate that the appropriate noise level depends on image resolution.

Interestingly, our empirical results suggest that the optimal fixed noise levels, s=0.8 for ImageNet (256x256) and s=0.35 for CIFAR-10 (32x32), both correspond to an effective log SNR close to zero when accounting for resolution. Specifically, under our noising formulation xs=(1−s)x0+sεx_s = (1-s)x_0 + s\varepsilon, the resolution-based log SNR becomes logSNReff(s,d)=2log⁡1−ss+2log⁡d64log SNR_{eff}(s,d) = 2\log{\frac{1-s}{s}} + 2\log{\frac{d}{64}}, following Hoogeboom et al. (2023), with 64x64 as the reference resolution. Setting this quantity to zero yields s=0.8 for 256x256 and s=0.333 for 32x32, closely matching our empirical findings.

These observations suggest that the benefit of fixed-noise pretraining may come from concentrating the pretraining objective on an effective intermediate denoising regime, where the corrupted image retains meaningful structure while requiring substantial denoising. Rather than implying that fixed-noise pretraining is better, our results indicate that a carefully chosen fixed noise level can also be effective.

One thing to also note is that when we use the reweighted logit-normal pretrained model as a flow matching model to generate samples with 100-step Euler, FID=10.09 and FDr^5=31.79, underperforming poorly to its post-training counterpart. This result shows how effective the distributional post-stage is, demonstrating that an appropriate recipe (our method) can significantly enhance the performance of a generative model.

[W5 + Q2] Training compute accounting

I am not sure what is included in the reported EFLOPs. Does the calculation include online generation of both active and reference samples?

We clarify that our reported training EFLOPs already include the cost mentioned by the reviewer, including active and reference sample generation, feature extraction for real and generated images, and backpropagation through the frozen feature encoders to the generated images. The detailed breakdown of the training cost is provided below.

Component, EFLOPsB160L160
Pretraining, including perceptual loss76.8703175.3278
Active + reference generation, plus active generator backward6.557522.8711
Real-image feature extraction13.542313.5423
Active-image feature extraction6.77116.7711
Reference-image feature extraction6.77116.7711
Active encoder forward replay6.77116.7711
Encoder input-backward + counted OT operations6.89966.8996
Total124.1831238.9542

We apply the same FLOPs counting protocol consistently to our models and baselines, ensuring a fair training-compute comparison.

References

  • Vincent et al., “Extracting and Composing Robust Features with Denoising Autoencoders,” ICML 2008.
  • Hang et al., “Improved Noise Schedule for Diffusion Training,” ICCV 2025.
  • Choi et al., “Perception Prioritized Training of Diffusion Models,” CVPR 2022.
  • Hang et al., “Efficient Diffusion Training via Min-SNR Weighting Strategy,” ICCV 2023.
  • Hoogeboom et al., “simple diffusion: End-to-end diffusion for high resolution images,” ICML 2023.
  • Chen, “On the Importance of Noise Scheduling for Diffusion Models,” arXiv 2023.

Thank you for your hard work. Your additional experiments have provided useful evidence for the effectiveness of fixed-noise pretraining on ImageNet, and your detailed compute breakdown has addressed my concern about training compute accounting. However, I still have some questions about the generality of the fixed-noise pretraining claim, as well as concerns regarding the writing, framing, and literature attribution.

W1. As I said in my review, I did not question the value of discovering that these tools work efficiently in combination. My concern was, and remains, that the terminology and framing obscure the precise nature of the contribution.

First, my concern is what the term "Priming Models" refers to. I see the contribution primarily as an insight or a training technique rather than an entirely new model family, so I find the term confusing.

Second, to answer the question in the title, the evidence does not suggest that we are overthinking one-step generation. Every component of this recipe builds on existing tools and techniques, including denoising pretraining, LPIPS and ConvNeXt perceptual losses, multi-scale features from frozen encoders, and Sinkhorn OT distribution matching. The results demonstrate that this two-stage training recipe can be effective, but they do not establish that previous approaches to one-step generation were unnecessarily complicated.

Another source of confusion is that while the additional experiments on ImageNet demonstrate the effectiveness of fixed-noise pretraining, the best-performing text-to-image configuration and the world modeling experiments still use the reweighted logit-normal noise distribution. The paper should explicitly explain the motivation for using logit-normal noise in these tasks and clarify whether the claimed sufficiency of fixed-noise pretraining extends beyond ImageNet. Otherwise, the scope of the claims regarding fixed-noise pretraining should be reconsidered.

The framing of this paper should describe the contribution accurately and give due credit to prior work. In my view, a more accurate framing would be that reconstruction-based pretraining provides an effective initialization for feature-space distributional post-training, improving its efficiency and performance.

W2. Again, I agree that using fixed-noise denoising pretraining for downstream generative modeling serves a different purpose from earlier discriminative work. However, the paradigm of fixed-noise denoising pretraining followed by downstream fine-tuning has already been explored in Brempong et al. and Zaidi et al. Citing general denoising autoencoders does not replace discussing prior work that specifically established fixed-noise pretraining. Notably, your rebuttal contrasts your task with Brempong et al. and Zaidi et al., but does not state whether you will include them in the revision. At a minimum, the paper should explicitly cite and discuss its connection to those earlier studies and give them due credit.

I will reconsider my score based on how the authors address these remaining concerns.

References

[1] Brempong et al., “Decoder Denoising Pretraining for Semantic Segmentation,” arXiv:2205.11423, 2022. https://arxiv.org/abs/2205.11423

[2] Zaidi et al., “Pre-training via Denoising for Molecular Property Prediction,” arXiv:2206.00133, 2022. https://arxiv.org/abs/2206.00133

Reviewer_pMpBOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Priming Models, a two-stage framework for one-step generation. A generator first learns image reconstruction from corrupted inputs, then trains feature-space distribution matching under pure Gaussian inputs. The main implementation uses an OT objective with a generated-sample self-interaction correction. The central contribution is demonstrating that simple reconstruction pretraining can effectively initialize distributional training without a pretrained diffusion or flow teacher.

Validation covers class-conditional ImageNet generation, text-to-image generation, and world modeling. The pixel-space ImageNet model achieves a one-step FID of 1.19 with substantially reduced reported training compute relative to the compared baselines. Ablations examine pretraining noise levels, perceptual supervision, alternative reconstruction tasks, distributional objectives, and feature spaces.

Critical strengths
  1. A simple and well-supported training principle. The paper demonstrates that reconstruction at a single fixed noise level can provide an effective initialization for distributional generation training. Ablations across corruption levels, reconstruction tasks, and distributional objectives support the finding beyond one particular implementation. This provides a useful design principle for training one-step generators without first obtaining a strong diffusion or flow teacher.

  2. Substantial quality–compute improvements across applications. On ImageNet, the method reports FID 1.19 at 238.95 EFLOPs, compared with pixel MeanFlow-H’s 2.15 at 1968.54 EFLOPs, with both training stages included in its compute accounting. Additional results on text-to-image generation and world modeling strengthen the practical significance of the approach and demonstrate benefits beyond a single benchmark.

Critical weaknesses

None

Questions for authors

I would appreciate clarification on the following points:

  1. How does the quality–compute comparison change when including the FD-loss–fine-tuned pMF baseline excluded from Table 1?
  2. In text-to-image generation, do different noise seeds produce meaningfully diverse images for the same prompt? Was within-prompt diversity evaluated quantitatively, e.g., compared with MiniT2I?
  3. How sensitive are final performance and training stability to the hyperparams, e.g., OT regularization strengths (0.001/0.01) and their mixing weights (0.75/0.25)?
5 comments

We thank the reviewer for highlighting our contributions. Below, we provide the experimental results that the reviewer suggested.

How does the quality–compute comparison change when including the FD-loss–fine-tuned pMF baseline excluded from Table 1?

  • pMF-L + FD-SIM: FID=0.87, FDr5=2.40, EFLOPs=990.20 (821.75+168.45)
  • pMF-H + FD-SIM: FID=0.87, FDr5=2.17, EFLOPs=2255.24 (1968.54+286.70)
  • Priming-B: FID = 1.47, FDr5 = 12.71, 85.75 EFLOPs.
  • Priming-L: FID = 1.19, FDr5 = 7.10, 238.95 EFLOPs.

Both FD-tuned pMF variants achieve lower FID and FDr5 than Priming-L. Priming Models nevertheless use much less compute (pMF-L 990.20 EFLOPs, while Priming-L 238.95 EFLOPs, about one quarter of the training compute)

We provide two details to help interpret this comparison.

First, FD-SIM directly minimizes Fréchet distances in the Inception-V3, SigLIP, and ViT-MAE feature spaces. FID uses Inception-V3, and FDr5 includes all three encoders. These scores therefore partly reflect performance in the feature spaces used for fine-tuning. In Table 1, all five evaluation encoders are held out during training.

Second, we take a complementary view of FD-loss as one possible objective for distributional post-training. As reported in Section 4.2 and Appendix E, Priming-L reaches an FID of 0.94 after 160 epochs of reconstruction pretraining and 20 epochs of FD post-training, narrowing the gap to FD-tuned pMF-L (FID 0.87). Our contribution is to show that reconstruction pretraining accelerates distributional post-training with existing objectives (not on developing new distributional objectives, to be clear).

In text-to-image generation, do different noise seeds produce meaningfully diverse images for the same prompt? Was within-prompt diversity evaluated quantitatively, e.g., compared with MiniT2I?

Appendix D reports precision and recall on ImageNet. To compare within-prompt diversity in text-to-image generation, we evaluated Priming and the multi-step MiniT2I baselines on 100 prompts, generating 16 images per prompt with different noise seeds.

For each prompt, we computed the mean pairwise cosine distance between DINOv2 embeddings and the mean pairwise LPIPS distance, then averaged over prompts.

  • JiT-100step (cfg=5.0): dinov2=0.10, lpips=0.5456
  • MF-4step: dinov2=0.133, lpips=0.57
  • Priming-1step: dinov2=0.1516, lpips=0.5045

Higher values indicate greater variation across seeds. Among the three models, Priming has the highest mean DINOv2 distance and the lowest mean LPIPS distance. Thus, Priming produces diverse outputs for a given prompt, with its ranking relative to the multi-step baselines depending on the metric.

How sensitive are final performance and training stability to the hyperparams, e.g., OT regularization strengths (0.001/0.01) and their mixing weights (0.75/0.25)?

Starting from the same 160-epoch Priming-B reconstruction checkpoint, we varied the weights assigned to the two OT regularization strengths during post-training. The weights below are listed in the order ε = 0.001, ε = 0.01.

OT configurationFID ↓FDr5 ↓
Baseline 75/251.38411.294
Pure ε=0.0011.37111.533
Pure ε=0.011.39411.162
Mixture 25/751.36811.321

All four configurations reached similar final scores. These results suggest that final performance is largely insensitive to the mixing weights over the range tested, showing the robustness of the Priming Model framework.

Thank you for the additional experiments. The FD-tuned pMF comparison clarifies the quality–compute trade-off, and the diversity results are helpful, though the ranking depends on the metric. I have two follow-up questions:

  1. Could you report the total training compute for Priming-L at FID 0.94, including both reconstruction pretraining and FD post-training? This would allow a more direct comparison with pMF-L + FD-SIM (FID 0.87 at 990.20 EFLOPs).

  2. Could you show uncurated grids of all 16 outputs for a few shared prompts, with Priming and JiT-100step side by side? This would help show whether changing the seed affects composition and viewpoint or mainly changes finer details.

We thank the reviewer for the prompt follow-up questions.

Training compute. Priming-L at FID 0.94 uses 175.3 EFLOPs for reconstruction pretraining (Priming-L-160) and 33.0 EFLOPs for 20 epochs of FD-SIM post-training, for a total of 208.3 EFLOPs. This is approximately 4.75× less training compute than pMF-L + FD-SIM (990.20 EFLOPs), with FID 0.94 versus 0.87.

Uncurated seed grids. We added six shared prompts on pages 1–6 of the supplementary PDF. For each prompt, the grids show all 16 outputs from Priming (one step) and JiT (100 steps, CFG 5.0) side by side, aligned by seed, with no output filtering.

We hope these address the reviewer's concern appropriately.

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

This paper proposes Priming Models, a two-stage approach to one-step generation. The network first learns to reconstruct corrupted images, then undergoes feature-space distributional training from Gaussian inputs. The fixed-noise reconstruction provides a useful initialization. The experiments in the paper span across ImageNet, text-to-image generation, and world modeling, with large improvements in the quality–compute trade-off over strong baselines.

Critical strengths
  1. The paper shows that distributional training can start from a reconstruction model with poor initial one-step samples. Experiments with different reconstruction tasks and distributional losses also support this finding beyond one particular configuration.

  2. The main result is very strong. Priming-L achieves ImageNet FID 1.19 at 238.95 EFLOPs, compared with 1.61 at 1476.39 EFLOPs for Drifting-L. Improvements are also shown in five evaluation encoders unused in the main training setup.

Critical weaknesses
  1. The Stage 2 in the paper seems to have a substantial overlap with W-Flow. Both use the same underlying Sinkhorn update, including independent generated batches for self-transport. Under matched settings, their core unguided gradients agree up to a constant factor. This overlap, however, is not clearly explained in the paper. The main contribution therefore appears to be reconstruction initialization and its empirical benefits, which should be more clearly distinguished from the existing distributional training method.

  2. It remains unclear how much of the gain in efficiency comes from the reconstruction part. Could an initial OT training phase with a larger learning rate and smaller batches and fewer inner iterations, provide comparable benefits? A comparison under the same total compute would help distinguish the effect of reconstruction from that of another method. The noise-level ablations support the chosen setting empirically, but it's unclear when fixed-noise reconstruction should be preferred or how its noise level should be selected.

Questions for authors

Could the authors address the weaknesses above, particularly the relationship between Stage 2 and W-Flow, and whether a tuned initial OT training phase can offer similar benefits to reconstruction pretraining?

6 comments

First of all, we appreciate the reviewer’s overall positive evaluation of our paper.

Comparison to W-Flow

Could the authors address the weaknesses above, particularly the relationship between Stage 2 and W-Flow

We agree that, in terms of the training objective, our Stage 2, instantiated with OT, is closely related to W-Flow’s, as discussed in Appendix A. Our OT objective and W-Flow’s velocity-based objective on the feature space have the same gradient up to a constant factor.

However, the two approaches differ in their motivation and formulation. W-Flow is derived from the perspective of Wasserstein gradient flows, where a velocity field is constructed to evolve the generated distribution toward the data distribution. In contrast, our formulation directly optimizes a distributional discrepancy using a gradient surrogate objective, without introducing the velocity field. Although these formulations lead to equivalent gradients up to a constant, our approach provides a direct distributional optimization perspective that does not rely on the gradient flow framework.

Most importantly, as the reviewer notes, the central distinction is how the generator is prepared for distributional optimization. W-Flow applies distributional training from scratch, whereas Priming first undergoes reconstruction pretraining. Our contribution is therefore to show that this simple preparation substantially improves the compute-quality trade-off, making distributional optimization much more effective.

Priming’s training recipe also dispenses with several additional components used by W-Flow, which might have helped address the difficulty of distributional training from random initialization. We discuss these differences in Appendix A and summarize them here for clarity:

  • Priming does not use learned style embeddings.
  • Priming does not use classifier-free guidance and therefore has no guidance scale to tune
  • Priming operates in pixel space, whereas W-Flow’s reported ImageNet models operate in VAE latent space.

The empirical gap is indeed striking: On ImageNet 256 × 256, Priming-B reaches FID 1.47 (85.75 EFLOPs), compared with FID 1.52 (2,030 EFLOPs) for W-Flow-B/2. Priming-L reaches FID 1.19 (238.95 EFLOPs), compared with FID 1.35 (2,840 EFLOPs) for W-Flow-L/2. Priming achieves lower FID with approximately 1/24 of W-Flow’s training compute at B scale and 1/12 at L scale.

OT training initialization instead of the reconstruction pretraining

and whether a tuned initial OT training phase can offer similar benefits to reconstruction pretraining?

To further demonstrate the reconstruction stage’s role, we tested the effect of reconstruction pretraining by comparing three Priming-B runs, each with the same budget of 85.75 EFLOPs. In particular, we tested OT training from scratch, and OT training with a larger learning rate (10^-5 to 10^-4, 10 times larger) and smaller batch size (1024 to 256).

  • Reconstruction pretraining (100k updates, 38.44 EFLOPs) → OT post-training (25k updates, 47.31 EFLOPs). FID = 1.47.
  • OT warm-up (batch size 256, learning rate 1e-4, 81k updates, 38.44 EFLOPs) → OT post-training (25k updates, 47.31 EFLOPs). FID = 138.94.
  • OT training from scratch (45k updates, 85.75 EFLOPs). FID = 203.8.

Taken together, our matched-compute experiment and the results in Table 1 show that reconstruction pretraining substantially improves the compute-quality trade-off.

Empirical choice of choosing the noise level

it's unclear when fixed-noise reconstruction should be preferred or how its noise level should be selected.

Per the reviewer’s suggestion, we investigated how the noise level could be selected a priori. Our empirical findings suggest that the selection of an effective noise level is related to the signal-to-noise ratio (SNR), while accounting for image resolution.

Prior works have highlighted the importance of noise-level selection in diffusion model training (Choi et al., 2022; Hang et al., 2023, 2025). In particular, Hang et al. (2025) show that emphasizing noise levels around log SNR = 0 can improve training efficiency. Furthermore, Hoogeboom et al. (2023) and Chen (2023) demonstrate that the appropriate noise level depends on image resolution.

In our Priming Model experiments, the best-performing fixed noise level was s=0.8 for ImageNet (256x256) and s=0.35 for CIFAR-10 (32x32) (the CIFAR-10 result was obtained using an independent implementation).

Noise level (s)0.150.250.350.450.550.650.700.750.800.84
FID ↓4.143.253.123.253.574.485.236.277.8810.26

Under our noising formulation xs=(1−s)x0+sεx_s = (1-s)x_0 + s\varepsilon, the resolution-based log SNR becomes logSNReff(s,d)=2log⁡1−ss+2log⁡d64log SNR_{eff}(s,d) = 2\log{\frac{1-s}{s}} + 2\log{\frac{d}{64}}, following Hoogeboom et al. (2023), with 64x64 as the reference resolution.

Both empirically optimal noise levels correspond to an effective log SNR close to zero. (According to the formula, the optimal noise level is s=0.8 for 256x256 and s=0.333 for 32x32.) This suggests that reconstruction pretraining benefits from an intermediate noise regime, where the image retains meaningful structure while requiring substantial denoising.

These observations highlight that the optimal noise level for Priming could have a connection to the SNR, providing an affirmative answer to the reviewer’s question.

References

  • Hang et al., “Improved Noise Schedule for Diffusion Training,” ICCV 2025.
  • Choi et al., “Perception Prioritized Training of Diffusion Models,” CVPR 2022.
  • Hang et al., “Efficient Diffusion Training via Min-SNR Weighting Strategy,” ICCV 2023.
  • Hoogeboom et al., “simple diffusion: End-to-end diffusion for high resolution images,” ICML 2023.
  • Chen, “On the Importance of Noise Scheduling for Diffusion Models,” arXiv 2023.

Thank you for your response and the additional experiments. However, my concerns are not fully addressed.

My first remaining concern is the characterization of Stage 2. A different motivation or derivation does not change the fact that its OT instantiation yields gradients equivalent to W-Flow’s under matched unguided settings, up to an overall scaling factor. This scaling does not establish a distinct core update rule. Figure 1 also shows that reconstruction pretraining alone produces poor one-step samples, with FID around 300, so the subsequent distributional training is essential to the final performance. Its relationship to existing work therefore warrants explicit acknowledgment where Stage 2 is introduced. Reconstruction pretraining itself is not a novel method, as the objective is very similar to a diffusion denoising loss.

The additional experiment highlights another issue: controlling total EFLOPs alone is not enough to isolate the factors that affect the generation quality and training compute trade-off. In the ablation setting reported in the W-Flow paper, W-Flow-B achieves an FID of 7.08. Based on the W-Flow-B/L compute figures reported in your rebuttal (2,030 and 2,840 EFLOPs), I estimate the compute of this ablation run to be approximately 45 EFLOPs, using a simple scaling model that accounts for the differences in training steps, batch size, and feature encoder width. Its FID is far better than those reported for the OT-only baselines in the rebuttal. This suggests that, in addition to total compute, the generation space, feature encoders, style embeddings, network depth and width, and patch size also matter when interpreting the trade-off between training compute and generation quality. Therefore, the comparison reported in the paper is not convincing enough to attribute the efficiency gains specifically to reconstruction pretraining, as it compares two sets of experiments under very different training settings without controlling these other factors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.