acceptodds
Under review as a conference paper at ICLR 2027

Unlocking Few-Step Diffusion for Faithful Previews

Abstract

Sampling latency compounds in diffusion workflows, where users generate and discard many candidates before keeping one. Surprisingly, the poor outputs of standard few-step samplers do not reflect a lack of reconstruction capacity: by optimizing only the initial noise, frozen 3–4-step samplers can closely reproduce their corresponding full-step outputs. Building on this finding, we learn corrections to the initial noise and denoising updates using endpoint supervision, improving correspondence with full-step outputs generated from the same noise and prompt. The resulting previews allow users to screen candidates cheaply and reserve full-step generation for promising ones. Input correction also transfers across sampling budgets without retraining. Experiments show substantial improvements in reference fidelity, including 53–78% lower reconstruction MSE than retrained LD3 on unconditional benchmarks, alongside improved ranking preservation and candidate selection on SD1.5, SDXL, and FLUX.1-dev. Code and model checkpoints are available at https://anonymous.4open.science/r/faithful-previews-anonymous-F2EB/.

Primary area:
generative models
Keywords:
diffusion model, preview, fast 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_1PohOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a novel approach to achieving few-step previews that closely match the results of full-step denoising by directly predicting the noise residual to select a better starting point. This approach can help users quickly identify better images or perform best-of-N selection for test-time inference scaling. Experimental results on SDXL and SD1.5 demonstrate that the proposed method outperforms previous approaches across different benchmarks, achieving lower MSE and higher PSNR.

Critical strengths
  • The key strength of this paper is its novel approach to updating the noise rather than designing a new scheduler. This makes the noise corrector simple yet effective. The quantitative results (Tables 1 and 2) demonstrate strong performance on toy examples, outperforming previous approaches. The qualitative results (Figure 11) also show promising outcomes, further supporting the effectiveness of the proposed method.
Critical weaknesses
  • One important weakness is that the trajectory correction is mainly motivated by empirical observations, with limited theoretical explanation. It would be helpful to provide more insight into how this correction affects the denoising process and why it is effective. For instance, is there any connection to consistency models or existing scheduler designs, but with a simpler formulation? Such an explanation could strengthen the overall contribution and provide more insight into the proposed method. A similar concern applies to the transfer step, which also seems to be an empirical design choice (Appendix C.3). In particular, it is unclear why the squared term helps improve the norm. Further explanation or justification would make this design more convincing.
  • Another concern regarding trajectory correction is whether it introduces additional training difficulties, as it may require backpropagation through multiple denoising steps, potentially making training more expensive or challenging for users. It would be helpful to report the training resources and computational costs, or clarify whether this is actually a concern in the proposed method. If so, discussing possible ways to mitigate this issue would further strengthen the paper.
Questions for authors
  • It would be helpful if the authors could provide a more mathematical explanation of how the proposed method relates to the denoising trajectory. This does not necessarily require developing a new theory, but establishing a connection to existing approaches, such as consistency models or scheduler designs, could provide more insight into why the method works.
  • Could the authors provide additional details on the training resources and computational time? It would also be helpful to clarify whether trajectory correction or input correction requires backpropagation through multiple denoising steps, and whether this introduces any training difficulties or instability. If so, how are these challenges addressed?
3 comments

Thank you for the thoughtful review and for recognizing the simplicity and effectiveness of our approach!

Summary of changes: We added a mathematical decomposition of trajectory error and a supporting SD1.5 experiment on per-step consistency. We also add information on the training cost.

The new SD1.5 analysis and reported cost are highlighted in blue in the revised paper, with their locations specified below.

1. Mathematical explanation of trajectory correction

Great point! We fully agree that our method would benefit from a deeper theoretical understanding.

Following your suggestion, we added a per-step consistency analysis and supporting experiments to address two questions: (1) Does LoRA correction make each student step more consistent with the teacher? (2) How does each corrected step affect the final reconstruction error, individually and through interactions with other steps?

Our starting point is Proposition A.1, which decomposes the squared L2 reconstruction error between student and teacher into per-step defect energies and their cross-step interactions (E=S+CE=S+C). A step's defect measures how much taking that student step changes the final output obtained by teacher continuation. This isolates each step's effect and shows how the effects combine.

We then measure these quantities on SD1.5 using 32 prompt–seed pairs, comparing our three-step LoRA-corrected student with the raw three-step DDIM baseline. The results are summarized below.

QuantityRGBLatent
Per-step defect-energy reduction vs. raw DDIM (steps 1/2/3)16.0% / 17.2% / 31.4%22.0% / 33.1% / 43.0%
LoRA defect energy offset by cross-step cancellation66.0%59.0%
Total MSE reduction vs. raw DDIM68.9%74.8%

These results show two findings. First, LoRA indeed makes every step more consistent with teacher continuation. Our method thus aligns with the consistency model’s principle of preserving the teacher-predicted endpoint along the sampling trajectory. Second, interestingly, the remaining defects partially cancel across steps, further reducing the final reconstruction error. The decomposition shows how smaller per-step defects and their partial cancellation contribute to the final error reduction.

Please see Section 3.3 and Appendix A.5 (blue text), including Proposition A.1, Figure 9, and Tables 6–7.

Step-count transfer. For input correction, we expect a smaller correction at larger sampling budgets because raw samples are typically closer to the full-step reference. The power term controls this decay; normalization preserves the original noise norm. Experimentally, stronger decay (p=2.5p=2.5) gives higher PSNR, while weaker decay (p=1.7p=1.7) gives higher HPS. The cause of this tradeoff requires further study. Please see Section 2 and Appendices A.4 and C.2 (Table 8).

2. Backpropagation and training requirements

Both methods backpropagate through the few-step student (rather than the teacher model) during training, updating the input corrector or step-specific LoRA adapters.

Training cost: The SD1.5/SDXL trajectory models use 8,192 updates on 4,000 prompts at five CFG scales. Training takes about 30.3 L40S GPU-hours for SD1.5 and 86.2 GPU-hours for SDXL.

FLUX uses 36,864 paired inputs and 1,548 updates, with gradient clipping and EMA. The selected EMA checkpoint is at update 1,280, with a measured training-loop cost of about 28.66 GH200 GPU-hours.

We report these numbers in Appendix B.3 of the revised manuscript.

Training stability and memory. We did not observe training instability in our experiments. We added a small profiling experiment to measure training memory and runtime on a single GH200 GPU with 95.1 GiB of visible memory. At 512×512, rank 64, and microbatch 1, our four-step FLUX student uses 37.7 GiB peak allocated memory and 3.02 s per training update. The same setup supports 16-step training updates with 61.1 GiB peak allocated memory, while 32-step training runs out of memory. Thus, our current few-step setup fits within the available GPU memory. To reduce backward computation, one can potentially detach the denoiser input.

We hope these additions make the connection between individual denoising updates and final reconstruction fidelity clearer. Thank you again for the helpful suggestions!

I appreciate the authors' response and the additional theoretical analysis. I maintain a positive outlook on the paper.

My final suggestion is to adopt a more modern model that is more widely used in real-world applications today, as SD1.5 and SDXL are less commonly utilized now.

We agree, and thank you for your positive review and helpful suggestion! Your feedback has helped us substantially strengthen the paper.

Reviewer_Ay2eOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to make few-step diffusion samples faithful previews of the full-step images. Then users can screen candidates cheaply and run full generation only for the ones they choose. The proposed method learns two corrections, both trained to match the full-step output from the same noise, including an input noise corrector and a trajectory correction. Experiments are conducted using Stable Diffusion and FLUX with standard datasets and metrics.

Critical strengths
  1. The motivation is good for a practical problem. In real use of image generation models, users often generate many candidates and discard most of them. Cheap previews that faithfully predict the full-step result address a real cost in this workflow.

  2. The method is standard, simple and effective. The approach applies standard techniques, such as noise correction and adapters. This is an interesting application of existing techniques. Transferring the input corrector to other step counts is a nice practical property.

Critical weaknesses
  1. The experiments may not be fully convincing yet. The previewing is a human-in-the-loop use case, but all evaluation relies on automatic preference scorers or an LLM agent. User studies are needed to show this is really useful for humans, especially since this is an application-focused paper.

  2. The evaluation should focus on text-conditioned generation, which is the motivating use case. Some of the experiments are only on unconditional models, and the abstract’s results also come from unconditional benchmarks.

  3. In addition, the number of prompts is relatively small.

  4. The writing can be improved. Section 3 largely restates Section 2 in more detail, mixing method description with experiments, so the reader goes through the same ideas twice. The readability is low and sometimes confusing.

Questions for authors

See above.

3 comments

Thank you for recognizing the practical motivation and simplicity of our approach!

Summary of changes: We added an SDXL human study, doubled the SD1.5/SDXL evaluation at CFG scales 1, 3, 5, and 7, and added a larger FLUX selection comparison. We revised the Abstract and Introduction to emphasize text-to-image previews, brought the FLUX comparison and human results into the main paper, and streamlined Sections 2–3.

Revised text is highlighted in blue, and we give the corresponding manuscript locations below.

1. Human evaluation

We agree that human evaluation is very important. We added an SDXL study in which users selected one of eight previews, then rated how well its corresponding full-step image matched their expectations (1–5). Across 45 prompts and 39 survey responses, we obtained 195 ratings per method. We will do our best to recruit more participants and extend this evaluation.

MethodMean ± SDMedian5/5 ratings
Ours4.29 ± 0.96554.4%
LD33.42 ± 1.30328.2%
DPM++3.27 ± 1.34324.6%

These results show that our previews help users anticipate the final images they select.

Please see Section 4.3 (Table 5 and Figure 8) and Appendix C.8 (Table 15) for the protocol and complete counts.

2. Greater emphasis on text-to-image generation

We revised the Abstract and Introduction to emphasize the text-to-image preview workflow. The main paper now includes the FLUX visual comparison and human study. Figure 7 illustrates the practical benefit: our previews better preserve the composition and appearance of the full-step reference.

Meanwhile, we have placed greater emphasis on text-to-image generation in the revised manuscript. Our new experiments include an SDXL human study, larger prompt evaluations on SD1.5, SDXL, and FLUX, and additional FLUX reconstruction and selection comparisons. We also added a theoretical analysis and an SD1.5 mechanism study to better understand trajectory correction and per-step consistency.

Please refer to the Abstract, Introduction, Sections 3.3 and 4.2–4.3, and Appendices A.5 and C.5.

We hope these additions make the practical value of our method for text-to-image previews clearer. Thanks!

3. Larger prompt evaluations

We doubled the reconstruction evaluation from 100 to 200 held-out COCO prompts for both SD1.5 and SDXL at CFG scales 1, 3, 5, and 7, with eight noises per prompt. Our method leads all four reported fidelity metrics on this expanded set. Please see Section 3.3 (Table 2) and Appendix C.1.

We added a separate FLUX selection comparison on 553 prompts, with eight candidates per prompt. Our four-step previews improve selected teacher-image ImageReward by 0.1177 over four-step SCFM. Please refer to Appendix C.5 (Table 12).

We will continue to collect more prompts to broaden and strengthen this evaluation.

4. Clearer organization

Thanks for the suggestion! We consolidated the shared objective, correction parameterizations, and preview–select–render workflow in Section 2. Section 3 now focuses on the empirical evidence for reconstruction and correction, while Section 4 presents the preview applications. Detailed formulations and training settings are collected in Appendices A–B. We hope these changes reduce redundancy and make the paper easier to follow.

We hope the human study and expanded text-to-image evaluation make the practical value of faithful previews clearer. We'd love to hear your thoughts

Thank you, and I have raised my score to 3: Weak acceptance.

For the user study, I understand that it is highly simplified and premature given the limited rebuttal time. It must be more rigorous and comprehensive to be reliably included in the later version of the paper. The user study should also investigate whether the method can actually improve productivity (not only a match rating) as claimed in the motivation.

For other experimental issues and writing issues, they also need to be refined in the later version of the paper.

Thank you for your positive assessment and constructive feedback! We will work to further strengthen the evaluation and improve the writing in the next version. We appreciate your helpful suggestions to improve our paper!

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

This paper propose a simple framework to train a preview model, basically with endpoint regression and backpropagate through denoising chain. It provides two implementations, one is optimizing the input initial noise by training a noise corrector, and the other is optimizing the entire trajectory by training a LoRA applied to each denoising step. The results show that the method has higher reference fidelity on SD1.5, SDXL, FLUX.1-dev compared to baselines.

Critical strengths
  • This paper deals with a practical and important problem of faithful and cheap preview of multi-step diffusion models, with a very simple method. The simplicity of the method and it's high performance is key strength of the paper.
  • Motivating experiment to optimize initial noise to reconstruct the teacher sample is interesting
  • Fig 7 (comparison on naive few step sampling, a few-step model (FLUX.1-schnell), proposed method, teacher sample (GT)) is very convincing. I think it should be in the main paper.
  • The paper shows the ranking correlation and regression, not only reconstruction-based metrics like MSE. This suggests the proposed method can be useful in practice
Critical weaknesses
  • I wonder the position of "preview" task compared to few-step distillation. Although not all the few-step objective enforces endpoint matching (for example, DMD only match teacher and student samples in distribution level), some representational subset of works indeed match the endpoint as well (for example, flow-map family like MeanFlow and Shortcut models. Also progressive distillation is one of the work). If these endpoint few step modeling is perfect, it can be directly used for preview (or even for generation). To claim some advantage over them, the paper should show either the inference efficiency (e.g., the proposed method can achieve better reference fidelity with lower cost), training efficiency (the proposed method can be trained very quickly), or better reference fidelity or better fidelity-cost tradeoff, something like that. If not, people can just use other few-step distillation methods.
  • I wonder if the proposed method is better than other paired-distillation baseline. It's not clear if the gain comes from the correction design (LoRA) or paired endpoint supervision. Authors could try other baselines like progressive distillation, direct endpoint supervision without backpropagagtion through denoising chain, etc)
  • The input corrector experiments are done in very small scale datasets and models, which does not mean much. I think the authors should provide one of SD1.5/SDXL plus one more recent model like FLUX in main paper, basically sharing the same setting between input correction and trajectory correction.
  • Cross-backbone related argument can be better clarified in terms of interpretation, since it anyway uses CelebA teacher target (on Bedroom backbone), so it's not unseen-domain generalization.
Questions for authors
  • What's the distinction of the proposed method compared to paired endpoint distillation? Is it about step-wise LoRA? endpoint loss? backpropagation through the denoising chain? How does the proposed method perform compared to other teacher-aligned distillation with the same architectural design (stepwise LoRA, etc)?
  • Could the authors provide comparison to shared LoRA student?
  • Could the authors provide comparison on full-chain gradient versus detaching the gradient of the denoiser? (approximating Jacobian of denoiser with an identity matrix)
  • What is the main claim on the advantage of the proposed method? Better reference fidelity given the same inference cost? or training cost? Could authors clarify the advantage and also show cost-fidelity tradeoff? It would substantially strengthen the paper.
2 comments

Thank you for recognizing the practical value of faithful previews and the simplicity of our approach!

Summary of changes: We added LoRA and gradient ablations, an SD1.5 mechanism study, comparisons with Progressive Distillation and SCFM, and a FLUX input-versus-trajectory comparison. We clarified the contribution and cross-backbone interpretation, and moved the FLUX visual comparison into the main paper.

We first present the new experiments, then explain how they support our contribution.

1. Additional experiments

1.1 Ablations

(a) Shared vs step-specific LoRA. We added an ablation comparing a shared adapter with separate adapters at each of the three denoising steps. We evaluate shared rank 8, shared rank 24, and step-specific rank 8; shared rank 24 matches the total parameter count of the three step-specific rank-8 adapters. Please see Appendix C.6 (Table 13, blue text).

LoRA designCIFAR-10 PSNR ↑CelebA-HQ PSNR ↑
Shared rank 829.040 ± 0.02628.239 ± 0.024
Shared rank 2429.284 ± 0.00928.648 ± 0.027
Step-specific rank 829.943 ± 0.02329.330 ± 0.006

(b) Full vs detached gradients. We compare full gradients with detaching the denoiser input. Please see Appendix C.6 (Table 14, blue text).

Gradient variantCIFAR-10 PSNR ↑CelebA-HQ PSNR ↑
Detached denoiser input29.526 ± 0.02128.238 ± 0.029
Full gradients29.943 ± 0.02329.330 ± 0.006

These ablations show that step-specific adaptation and full-chain gradients each improve fidelity under the same paired endpoint supervision.

We also added a mechanism study to better understand why our method is effective. Our experiments suggest that LoRA makes each step better preserve the teacher-predicted final output, and the remaining stepwise deviations partially cancel. Please refer to Appendix A.5 (Proposition A.1 and Tables 6–7).

1.2 Comparisons with distillation methods

Thank you for suggesting these comparisons. For unconditional generation, we added a comparison with Progressive Distillation (PD) on LSUN-Church. Our input correction improves PSNR over four-step PD by 8.77 dB, with a 23.4% LPIPS reduction. Please see Appendix C.9 (Table 16).

For text-to-image generation, we added comparisons with Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (SCFM; Cai et al., 2025), a shortcut-style distillation method on FLUX. Results are shown below.

MethodReconstruction PSNR ↑Selected-image ImageReward ↑Total selection NFE
SCFM, 4 steps14.4691.101760
SCFM, 8 steps15.3841.131492
Ours, 4 steps18.9021.219460

Notably, our four-step method outperforms eight-step SCFM in both reference fidelity and selected teacher-image ImageReward, with a lower selection budget (60 vs. 92 NFE). Please refer to Appendix C.5 (Tables 11–12) for details.

Nevertheless, the methods have different goals. Distillation focuses on few-step image quality. Eight-step SCFM scores higher on its own images but provides less faithful previews. The PD comparison also suggests that endpoint supervision alone may not ensure faithful previews.

1.3 Input versus trajectory correction on text-to-image models

The paper includes an SD1.5 input-corrector comparison, matching the data, endpoint supervision, and training-update budget. Trajectory correction improves PSNR from 20.60 to 23.25 dB, while using three backbone calls instead of four. Please see Section 3.2 and Appendix C.3 (Table 9).

Following your suggestion, we also added a FLUX comparison. Trajectory correction improves PSNR from 15.912 to 17.582 dB, using four backbone calls instead of five. This extends the comparison to a more recent text-to-image model. Please refer to Appendix C.3 (Table 9, blue text).

1/2

2. Contribution and practical advantage

Our main contribution is a simple method for learning fast, faithful few-step previews. More precisely:

(1) Better preview performance. We achieve higher reference fidelity than the evaluated distilled models in both conditional and unconditional generation. These include PD on LSUN-Church and FLUX.1-schnell and SCFM on FLUX. Our four-step model also outperforms eight-step SCFM in selected teacher-image ImageReward. These results demonstrate the practical benefit of training specifically for faithful previews. Please see 1.2 above and Appendix C.5.

(2) Simple and training-efficient. We learn input corrections or step-specific LoRA adapters while keeping the pretrained backbone frozen. Training uses endpoint supervision through the few-step student, with precomputed teacher outputs. Our SD1.5/SDXL models require 30.3/86.2 L40S GPU-hours, and FLUX uses 28.66 GH200 GPU-hours through the selected checkpoint. Our training costs are generally lower than these reported distillation budgets (albeit for a more specific task). For context, DMD reports approximately 2,592 A100 GPU-hours for SD1.5, and DMD2 approximately 3,840 A100 GPU-hours for SDXL. SCFM reports a comparable training cost for FLUX, but our method provides more faithful previews. This provides a practical route to faithful previews with modest training compute. Please refer to Appendix B.3 for our training settings and costs.

(3) A novel perspective on preview modeling. Our oracle experiments reveal that poor raw few-step outputs can hide substantial reconstruction capacity. Learning a reusable correction makes this capacity useful without per-image optimization at inference. Together with the ablations and mechanism study in 1.1, this establishes a correction-based preview paradigm and helps explain why such a simple approach can work.

Thank you for this insightful comment! We have revised the Introduction to make this contribution explicit and sharpen the paper's focus on fast, faithful previews. Please see the Introduction (blue text) and Sections 2–3 for the motivation, formulation, and supporting experiments. We hope this better conveys the contribution and practical value of our work.

3. Clarifications

Cross-backbone transfer. The corrector is trained through a Bedroom backbone using CelebA teacher targets, then applied to a CelebA-HQ backbone. The revised Section 3.4 explicitly identifies this as cross-backbone transfer with target-domain supervision. The experiment shows that the transferred correction improves teacher agreement and perceptual preference scores. Please see Section 3.4 (blue text) and Appendix C.4 (Table 10).

Main-paper visualization. We are happy to hear you find Figure 7 convincing! Following your suggestion, we have now moved the FLUX.1-schnell comparison to Section 4.2 (Figure 7), illustrating how our previews preserve the full-step reference's composition and appearance.

We hope these comparisons clarify the source of the gains and the practical value of faithful previews. Thank you again for the helpful suggestions!

2/2

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.