acceptodds
Under review as a conference paper at ICLR 2027

Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling

Abstract

Hardware-agnostic strategies for accelerating text-to-image diffusion, such as timestep distillation and feature caching, can reduce inference time without custom kernels or system-level optimization. Among them, multi-resolution generation strategies have recently received broad attention, achieving more than 5x speedup without training. However, the design of upsampling in the latent space, together with the selective modification of partial regions, causes these methods to exhibit noticeable blurring or artifacts. To this end, we propose MrFlow, a training-free multi-resolution acceleration strategy for pretrained flow-matching models built upon a staged low-to-high-resolution pipeline. MrFlow first rapidly generates the main structure at low resolution, then performs super-resolution in the pixel space using a lightweight pretrained GAN-based model, subsequently injects low-strength noise to enable high-frequency resampling, and finally refines the details at high resolution in one step while preserving structure. Quantitative and qualitative results on FLUX.1-dev and Qwen-Image show that MrFlow exploits quadratic token reduction and reduced step requirement of low-resolution sampling to achieve 10x end-to-end acceleration, while keeping the OneIG-En gap on Qwen-Image within 1%, significantly exceeding other training-free acceleration strategies, and requiring no training or runtime dynamic identification. MrFlow can also be directly combined orthogonally with pre-trained timestep distillation strategies, achieving even higher generation acceleration of up to 25x.

Primary area:
generative models
Keywords:
Generation Acceleration, Multi Resolution, Generative Model, Diffusion Model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gSxTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MrFlow is a training-free multi-resolution acceleration method for pretrained flow-matching T2I models (FLUX.1-dev, Qwen-Image). The paper argues that prior multi-resolution methods such as RALU, SPEED, and LSSGen upsample in the latent or frequency domain, which introduces blur and artifacts. It proposes a four-stage pipeline instead. First, it generates the global structure at low resolution in about 12 steps. Second, it decodes to pixel space and applies a pretrained Real-ESRGAN for ×2 super-resolution. Third, it re-encodes with the VAE and injects weak noise (σ≈0.1–0.15) so that erroneous high-frequency details from SR can be resampled. Fourth, it refines in a single high-resolution step. The authors justify the weak-noise design with the Wiener gain of a local Gaussian model: at a given noise level, the SR input is retained less along low-variance (high-frequency) directions. On Qwen-Image, MrFlow achieves 10.3× end-to-end speedup while keeping the OneIG-En gap within 1%, outperforming caching-based and other multi-resolution baselines at similar speeds. Combined with timestep-distilled models like Pi-Flow, it reaches up to 25× without any additional training.

Critical strengths
  1. The method is training-free, simple, and practical. It requires no runtime region identification or dynamic statistics and only combines an existing diffusion model, the VAE, and a pretrained SR network, which makes it easy to implement and reproduce. It applies directly to multiple backbones, including FLUX.1-dev, Qwen-Image, and FLUX.2 Klein (4B/9B), and the reported speedups are end-to-end measurements that include text encoding, VAE, and SR overhead.

  2. The quality-speed trade-off is strong in the aggressive acceleration regime. At around 9× on Qwen-Image, TeaCache (GenEval 0.26) and DB-Taylor (0.09) collapse, whereas MrFlow maintains 0.86 even at 10.3×. In the trade-off curves of Figure 5, MrFlow also lies consistently above native step reduction.

  3. The division of roles between stages is supported experimentally. The seed experiment showing roughly an 8000× difference in low-frequency MSE between varying the LR seed and the HR seed (Table 12), the band-wise differences between SR inputs and refined outputs (Table 13), and the sensitivity to low- and high-frequency perturbations (Table 9) consistently support the core claim that the LR stage determines structure while the HR stage refines details.

Critical weaknesses
  1. The novelty is limited. The pipeline of generating at low resolution, upscaling in pixel space, injecting weak noise, and denoising again at high resolution is essentially the same structure as Hires.fix (low-resolution generation followed by high-resolution repainting) and SDEdit, both of which the authors themselves cite in Appendix A. The differences are closer to configuration choices, namely the GAN-based upscaler, σ≈0.1, and 1-step refinement. Nevertheless, the main text does not clearly discuss how the method differs from these works, and there is no quantitative comparison with Hires.fix/SDEdit-style baselines under the same step budget. While the experiments are dense, I do not consider the methodology itself to be highly novel.

  2. The reliance on an additional pretrained network raises concerns about the method categorization and the fairness of comparisons. MrFlow uses a separately trained Real-ESRGAN as its upsampler. In contrast, the authors classify LSSGen as training-dependent because it trains a lightweight latent upsampler. In both cases, the method depends on an upsampler trained outside the diffusion model, and the only difference is who trained it. The categorization is therefore inconsistent. Furthermore, RALU and SPEED perform multi-resolution acceleration using only the capability of the pretrained diffusion model, so comparing them against MrFlow, which exploits an external SR prior, is not a like-for-like comparison. Notably, Table 7 shows that the interpolation-based MrFlow obtains the same quantitative scores as the Real-ESRGAN version, meaning the advantage on quantitative metrics is unrelated to the SR network. On the other hand, the qualitative comparisons (Figures 1, 7, 8), which serve as the main evidence of superiority over RALU and SPEED, appear to rely heavily on the detail prior of Real-ESRGAN. The authors should present the same qualitative comparisons using MrFlow with interpolation and no external network.

  3. The theoretical analysis of the noise strength lacks predictive power. The paper derives a lower bound on σ_t from a local Gaussian model and reports σ*≈0.075 from high-frequency band measurements. However, since the analysis provides only a lower bound, the values actually used (0.1–0.15) are ultimately chosen empirically, and the theory does not serve as a selection rule. Given the mid-frequency estimate of 0.16 in Table 8, the range 0.1–0.15 could be interpreted as one that resamples high frequencies while preserving mid frequencies, but the paper does not make this connection. In addition, the assumption that the eigen-directions of Λ can be replaced by latent FFT bands is not validated, and the gap between the analysis, which is based on the posterior mean (MMSE), and the actual refinement, which uses a single Euler step, is not addressed. Most importantly, there is no quantitative sweep of final quality over σ_t; Tables 9 and 10 use only a few σ values for diagnostic purposes. The analysis therefore appears to be a post-hoc explanation of a value that empirically works well. The authors should report quantitative metrics for σ ∈ {0.05, 0.075, 0.1, 0.15, 0.2, 0.3} or similar and show whether a performance transition actually occurs around σ*.

Questions for authors
  1. Could you provide quantitative and qualitative comparisons with Hires.fix/SDEdit-style baselines under the same step budget?
  2. Could you provide qualitative comparisons with RALU and SPEED using the interpolation-based MrFlow, without the external SR network?
  3. Could you report a quantitative sweep over σ_t and show whether a performance transition occurs around σ*?
10 comments
AuthorsOpenReview ↗

Response to Reviewer gSxT

Thank you for your careful review of our work. And we provide detailed responses to your comments as follows:

W1. The novelty is limited. The pipeline of generating at low resolution, upscaling in pixel space, injecting weak noise, and denoising again at high resolution is essentially the same structure as Hires.fix (low-resolution generation followed by high-resolution repainting) and SDEdit, both of which the authors themselves cite in Appendix A. The differences are closer to configuration choices, namely the GAN-based upscaler, σ≈0.1, and 1-step refinement. Nevertheless, the main text does not clearly discuss how the method differs from these works, and there is no quantitative comparison with Hires.fix/SDEdit-style baselines under the same step budget. While the experiments are dense, I do not consider the methodology itself to be highly novel.

A1. Thank you for the suggestion. A clearer comparative discussion will help readers understand MrFlow. Overall, as a method aimed at acceleration, MrFlow differs from these multi-resolution generation or noise-injection-based repainting strategies in application domain, underlying motivation, and pipeline design; its central idea is to assign the unavoidable multi-step denoising process to the low-resolution stage as much as possible, while the expensive high-resolution stage needs only a single step to complete the final refinement. We have briefly summarized this in the General Response, and elaborate below.

Differences from multi-resolution generation strategies such as Hires.fix:

  1. The application domains are entirely different. As stated in the related work (line 144) and the background section of the appendix (line 725) of our original manuscript, the above strategies aim to extrapolate beyond the recommended resolution natively supported by the model for inference, thereby raising the upper limit of the resolution the model can generate, with acceleration not being their central objective. MrFlow, in contrast, focuses entirely on generation acceleration, and the final generated images are at the model's recommended resolution, without generating beyond the resolution supported by the pretrained model.
  2. The underlying motivation and corresponding pipeline details are different:
    • Resolution configuration: The above strategies are motivated by the model's inability to directly construct global layouts in a higher-resolution space, so they first generate at the recommended resolution and then extrapolate to a higher resolution for refinement. The central motivation of MrFlow, however, is the observation that existing flow-matching-based diffusion models always retain complete global-structure generation capabilities even at the next lower resolution level, and can more easily be directly reduced to fewer steps; in the high-resolution space where the model excels, only high-frequency detail refinement is performed, rather than complete generation or reconstruction.
    • Upsampling: The above strategies do not emphasize or investigate differences in the effects of different upsampling methods. MrFlow points out the significant impact of different upsampling methods on generation quality and acceleration, explains in detail why pixel-space upsampling is chosen to connect resolutions based on measured results, and also explains why GAN-based Real-ESRGAN is chosen as the super-resolution model based on high- and low-frequency analysis.
    • Noise strength: The above strategies typically use a noise strength of no less than 0.5 after upsampling, significantly greater than the 0.15 advocated by MrFlow. MrFlow points out and demonstrates that super-resolved images differ from ideal images only in high-frequency details, which are precisely what Gaussian noise is good at overwriting and what a diffusion model is good at removing at the end of inference, closer to clean data; this property has not been pointed out in similar papers.
    • Step allocation across stages: The above strategies do not focus on the number of sampling steps at each stage, and in particular they typically use 10+ steps at the higher resolution reached through extrapolation. Specifically, as shown in Figure 10 of I-Max, text at the highest resolution is blurry with a 2-step configuration, and the number of steps has to be increased to 20 to become relatively stable. MrFlow, in contrast, explicitly points out and demonstrates that, in our pipeline, the low-resolution stage can easily be directly reduced to fewer steps (12 steps), and high-resolution refinement needs only 1 step to become stable, demonstrating the efficiency advantage of acceleration-oriented MrFlow when generating images at normal resolutions.
AuthorsOpenReview ↗

Response to Reviewer gSxT

Differences from noise-injection-based repainting strategies such as SDEdit:

  1. The task objectives are entirely different. As stated in the related work (line 155) and the background section of the appendix (line 755) of our original manuscript, there are typically large stylistic or structural differences between the inputs and outputs of SDEdit (as in style transfer or detail editing tasks). In MrFlow, adding noise and then denoising is intended to correct the erroneous high-frequency details left by the preceding fast low-resolution sampling and upsampling.
  2. The explanations of the underlying principles are different. Based on SDEs, SDEdit provides a strategy for connecting input and output distributions through noise injection, thereby transforming an input image into the target distribution. However, SDEdit cannot more quantitatively characterize the gap between the upsampled low-resolution distribution faced by MrFlow and the target high-resolution distribution. To address this, we conduct a new analysis based on high and low frequencies, and use statistics and calculations on real data to provide an estimate of a suitable noise strength that agrees with the experiments, providing a truly measurable and verifiable basis for explaining the mechanism of MrFlow.
  3. The pipeline details are different. To achieve editing, SDEdit typically relies on a relatively high noise strength, cannot be applied at low noise strengths, and also requires tens of sampling steps after noise injection. MrFlow, however, needs a noise strength of only 0.15 to achieve clearly perceptible single-step high-resolution refinement, with illustrative examples shown in Figures 3 and 4 of the main text.

Regarding the comparisons with Hires.fix and SDEdit under the same step budget that you mentioned, we conduct more detailed experiments in our subsequent response to your question Q1, A4. Hires.fix does not consider the need to accelerate generation at normal resolutions at all, nor does it consider the timestep-shift adjustments associated with few-step generation, and cannot achieve acceleration efficiency and generation quality comparable to MrFlow under the same total step budget. SDEdit considers neither the high similarity and high-/low-frequency characteristics between super-resolved and target images that are unique to MrFlow, nor the single-step denoising scenario unique to MrFlow; the related SDE theory is also redundant for flow matching and cannot guide noise injection and refinement in MrFlow.

We have briefly pointed out in both the main text and appendix of the original manuscript that MrFlow differs entirely from these strategies in motivation and application domain, among other aspects; the above provides a more detailed comparative supplement, and we have now added the relevant citations and discussion prominently in Section 2 and Appendix Section A.2 of the revised version.

AuthorsOpenReview ↗

Response to Reviewer gSxT

W2. The reliance on an additional pretrained network raises concerns about the method categorization and the fairness of comparisons. MrFlow uses a separately trained Real-ESRGAN as its upsampler. In contrast, the authors classify LSSGen as training-dependent because it trains a lightweight latent upsampler. In both cases, the method depends on an upsampler trained outside the diffusion model, and the only difference is who trained it. The categorization is therefore inconsistent. Furthermore, RALU and SPEED perform multi-resolution acceleration using only the capability of the pretrained diffusion model, so comparing them against MrFlow, which exploits an external SR prior, is not a like-for-like comparison. Notably, Table 7 shows that the interpolation-based MrFlow obtains the same quantitative scores as the Real-ESRGAN version, meaning the advantage on quantitative metrics is unrelated to the SR network. On the other hand, the qualitative comparisons (Figures 1, 7, 8), which serve as the main evidence of superiority over RALU and SPEED, appear to rely heavily on the detail prior of Real-ESRGAN. The authors should present the same qualitative comparisons using MrFlow with interpolation and no external network.

A2. Thank you for the suggestion. First, we provide a unified response: a significant difference between MrFlow and these training-free multi-resolution acceleration strategies is that LSSGen, RALU and SPEED all perform upsampling in latent space, whereas MrFlow chooses to perform upsampling in pixel space, particularly using GAN-based Real-ESRGAN; this choice is precisely one of MrFlow's core contributions. Specifically:

  1. Categorization of LSSGen: Because LSSGen trains a latent upsampler, and the VAEs used by different diffusion models are entirely different, meaning that their latent spaces are entirely different, it necessarily requires training different upsamplers for different diffusion models. MrFlow, regardless of which model it is applied to, performs upsampling in pixel space and does not depend on the choice of a specific model; moreover, it uses a very lightweight pretrained super-resolution network that has been open-source for over 4 years, rather than a super-resolution network specifically subjected to continued training or adaptation, so we believe that MrFlow is still training-free
  2. Fairness of comparisons with RALU and SPEED: We emphasize that choosing pixel-space super-resolution, using pretrained GAN-based Real-ESRGAN, and requiring a noise strength of only 0.15 afterward are themselves among MrFlow's distinctive observations and innovative contributions, and can be used without any additional continued training. Related works such as RALU and SPEED do not focus on this point and use higher noise strengths; the resulting larger number of high-resolution steps is precisely why they have difficulty achieving higher speedups than MrFlow. Therefore, although MrFlow introduces an additional upsampler, we believe that this does not affect its practicality as a training-free strategy
  3. The relationship between SR and quantitative metrics: This is indeed the case in the original manuscript's results, but there are clear differences in human visual perception, as shown in Figure 3. It may be that limitations of mainstream evaluation metrics such as GenEval and DPG-Bench prevent these differences from being measured quantitatively. Therefore, following Reviewer UhWG's suggestion, we have added results on metrics such as MUSIQ[1] and DeQA[2], as shown in Table 1 below. Note that while Real-ESRGAN outperforms Bicubic and SwinIR, it still falls short of OSEDiff. We believe that OSEDiff, which is trained based on a diffusion model, also demonstrates good compatibility with MrFlow, but given that neither its model size nor its pretraining scale can be considered lightweight, we include it only as one of the higher-potential options for MrFlow rather than as our main result.
Reviewer_UhWGOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose MrFlow, a training-free sampling pipeline to accelerate pretrained flow-matching text-to-image models. Specifically, there are four steps in the pipeline:

  1. Generate the image at half resolution with few denoising steps (usually 12)
  2. Decode with the VAE and upsample x2 in pixel space with a pretrained Real-ESRGAN(super resolution).
  3. Re-encode with the VAE and renoise with low-strength noise.
  4. Do one denoising step at full resolution to refine details.

On multi-step diffusion models, MrFlow significantly speedups the inference, with slightly lower quality.

Critical strengths
  1. The method is simple and can be plug and play to any image diffusion models. Training free characteristics make it easy to be applied by users.

  2. In comparison with Teacache, DB-Taylor, RALU and SPEED the performance on Qwen-Image doesn't degrade as the speed up ratio increases.

Critical weaknesses
  1. Novelty is limited, I believe the that generating at low resolution followed by pixel space upscaling and re-noising for high resolution refinement is quite a standard way. SDEdit shares the very similar idea. Even some ultra-high-res generation method like HiFlow and I-Max they both use this generate-upscale-refine idea.

  2. I believe the benchmarks cannot show the real quality changes, since those benchmarks are only in semantic level, the low-resolution results can already achieve good scores, low-resolution + bicubic interpolation also gets the same scores. Therefore, the HR refinement, shows no benefit in any quantitative metric. Please add more metrics for evaluation.

  3. Most of the current image generators have their corresponding "Turbo" version which is normally a 4 or 8 step model with similar or better performance than the base 28/50 step model. MrFlow shows no significant speedup for the distilled model, therefore, I believe it is not that useful in practice.

Questions for authors

Is this method able to generate higher resolutions?

I would also love to see that if this method is able to do well in 4 step distilled model like FLUX2-Klein-4B, where the time schedule is highly shifted to 1, 0.9674, 0.9081, 0.7672 → 0

13 comments
AuthorsOpenReview ↗

Response to Reviewer UhWG

Thank you for your careful review of our manuscript and your valuable suggestions. Our responses to your feedback are as follows:

W1. Novelty is limited, I believe the that generating at low resolution followed by pixel space upscaling and re-noising for high resolution refinement is quite a standard way. SDEdit shares the very similar idea. Even some ultra-high-res generation method like HiFlow and I-Max they both use this generate-upscale-refine idea.

A1. Thank you for the suggestion. A clearer comparative discussion will help readers understand MrFlow. Overall, as a method aimed at acceleration, MrFlow differs from these multi-resolution generation or noise-injection-based repainting strategies in application domain, underlying motivation, and pipeline design; its central idea is to assign the unavoidable multi-step denoising process to the low-resolution stage as much as possible, while the expensive high-resolution stage needs only a single step to complete the final refinement. We have briefly summarized this in the General Response, and elaborate below.

Differences from multi-resolution generation strategies such as HiFlow and I-Max:

  1. The application domains are entirely different. As stated in the related work (line 144) and the background section of the appendix (line 725) of our original manuscript, the above strategies aim to extrapolate beyond the recommended resolution natively supported by the model for inference, thereby raising the upper limit of the resolution the model can generate, with acceleration not being their central objective. MrFlow, in contrast, focuses entirely on generation acceleration, and the final generated images are at the model's recommended resolution, without generating beyond the resolution supported by the pretrained model.
  2. The underlying motivation and corresponding pipeline details are different:
    • Resolution configuration: The above strategies are motivated by the model's inability to directly construct global layouts in a higher-resolution space, so they first generate at the recommended resolution and then extrapolate to a higher resolution for refinement. The central motivation of MrFlow, however, is the observation that existing flow-matching-based diffusion models always retain complete global-structure generation capabilities even at the next lower resolution level, and can more easily be directly reduced to fewer steps; in the high-resolution space where the model excels, only high-frequency detail refinement is performed, rather than complete generation or reconstruction.
    • Upsampling: None of the above strategies emphasize or investigate differences in the effects of different upsampling methods. MrFlow points out the significant impact of different upsampling methods on generation quality and acceleration, explains in detail why pixel-space upsampling is chosen to connect resolutions based on measured results, and also explains why GAN-based Real-ESRGAN is chosen as the super-resolution model based on high- and low-frequency analysis.
    • Noise strength: The above strategies typically use a noise strength of no less than 0.5 after upsampling, significantly greater than the 0.15 advocated by MrFlow. MrFlow points out and demonstrates that super-resolved images differ from ideal images only in high-frequency details, which are precisely what Gaussian noise is good at overwriting and what a diffusion model is good at removing at the end of inference, closer to clean data; this property has not been pointed out in similar papers.
    • Step allocation across stages: None of the above strategies focus on the number of sampling steps at each stage, and in particular they typically use 10+ steps at the higher resolution reached through extrapolation. Specifically, as shown in Figure 10 of I-Max, text at the highest resolution is blurry with a 2-step configuration, and the number of steps has to be increased to 20 to become relatively stable. MrFlow, in contrast, explicitly points out and demonstrates that, in our pipeline, the low-resolution stage can easily be directly reduced to fewer steps (12 steps), and high-resolution refinement needs only 1 step to become stable, demonstrating the efficiency advantage of acceleration-oriented MrFlow when generating images at normal resolutions.
AuthorsOpenReview ↗

Response to Reviewer UhWG

Differences from noise-injection-based repainting strategies such as SDEdit:

  1. The task objectives are entirely different. As stated in the related work (line 155) and the background section of the appendix (line 755) of our original manuscript, there are typically large stylistic or structural differences between the inputs and outputs of SDEdit (as in style transfer or detail editing tasks). In MrFlow, adding noise and then denoising is intended to correct the erroneous high-frequency details left by the preceding fast low-resolution sampling and upsampling.
  2. The explanations of the underlying principles are different. Based on SDEs, SDEdit provides a strategy for connecting input and output distributions through noise injection, thereby transforming an input image into the target distribution. However, SDEdit cannot more quantitatively characterize the gap between the upsampled low-resolution distribution faced by MrFlow and the target high-resolution distribution. To address this, we conduct a new analysis based on high and low frequencies, and use statistics and calculations on real data to provide an estimate of a suitable noise strength that agrees with the experiments, providing a truly measurable and verifiable basis for explaining the mechanism of MrFlow.
  3. The pipeline details are different. To achieve editing, SDEdit typically relies on a relatively high noise strength, cannot be applied at low noise strengths, and also requires tens of sampling steps after noise injection. MrFlow, however, needs a noise strength of only 0.15 to achieve clearly perceptible single-step high-resolution refinement, with illustrative examples shown in Figures 3 and 4 of the main text.

We have already pointed out in the original manuscript that the motivations and application domains, among other aspects, are entirely different; the above provides a more detailed comparative supplement, and we have now added the relevant citations and discussion prominently in Section 2 and Appendix Section A.2 of the revised version.

AuthorsOpenReview ↗

Response to Reviewer UhWG

W2. I believe the benchmarks cannot show the real quality changes, since those benchmarks are only in semantic level, the low-resolution results can already achieve good scores, low-resolution + bicubic interpolation also gets the same scores. Therefore, the HR refinement, shows no benefit in any quantitative metric. Please add more metrics for evaluation.

A2. Thank you for your suggestion. As you noted, these most widely used evaluation metrics mostly rely on object detection or semantic alignment scores and do not sufficiently assess detail quality. Although they have already shown a trend toward relatively better overall results for MrFlow, they indeed do not fully reflect MrFlow's advantages in the visual appearance of details over other training-free strategies. To address this, we have added MUSIQ[1] and DeQA[2] results on Qwen-Image. MUSIQ and DeQA assess the perceptual quality of the image itself, complementing the original metrics' assessment of text–semantic alignment. Specifically, MUSIQ simultaneously analyzes local patches of the image at its original size and downscaled versions of the overall image, combining local details and overall appearance to predict human quality scores; its multi-scale design helps evaluate visual-quality issues such as blur and distortion. DeQA uses a multimodal model trained on human image-quality ratings to predict the probabilities that an image belongs to the quality levels “bad, poor, fair, good, excellent,” and then takes a weighted sum to obtain a continuous quality score; its evaluation target is image quality, rather than whether the objects or relationships in the prompt appear.

As shown in Tables 1 and 2 below, Tables 1 and 2 with the added metrics more comprehensively demonstrate MrFlow's quality advantage over other training-free acceleration strategies, with the best results at both the medium acceleration setting of around 5x and the high acceleration setting of around 10x, especially at the highest acceleration setting on Qwen-Image. Table 3 shows the consistent gains brought by single-step high-resolution refinement: with Real-ESRGAN, MUSIQ increases from 71.184 to 71.591, and DeQA increases from 4.1906 to 4.1998. Also note that while Real-ESRGAN outperforms Bicubic and SwinIR, it still falls short of OSEDiff. We believe that OSEDiff, which is trained based on a diffusion model, also demonstrates good compatibility with MrFlow, but given that neither its model size nor its pretraining scale can be considered lightweight, we include it only as one of the higher-potential options for MrFlow rather than as our main result.

We have incorporated the test results with these expanded metrics into Section 4 (Tables 1–3) of the revised paper and evaluated them in all of our experimental responses in this rebuttal, and have further added comprehensive results on FLUX and few-step models to Section 4 (Tables 1 and 2) and Appendix Sections D.1 and D.2 of the revised manuscript. Metric configurations are described in Section 4.2 and Appendix B of the revised manuscript. Thank you again for your suggestion, which has provided more objective support for demonstrating the strong performance of MrFlow.

Reviewer_x744OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces MrFlow, a training-free, multi-resolution acceleration strategy designed for pre-trained Flow Matching models. The method employs a multi-stage pipeline—consisting of low-resolution structural generation, pixel-level super-resolution, high-frequency resampling with low-intensity noise injection, and single-step high-resolution detail refinement—to accelerate the inference process. Experiments conducted on FLUX.1-dev and Qwen-Image demonstrate an end-to-end speedup of up to approximately 10x, achieved without the need for additional training.

Critical strengths
  • Exceptional Training-Free Efficiency: MrFlow accelerates the generation speed of modern large-scale models (such as FLUX.1-dev and Qwen-Image) by approximately 10x—without requiring any fine-tuning or complex, dynamic runtime identification—while effectively preserving both image quality and semantic alignment.

  • This staged design is motivated by a clear rationale: the low-resolution stage handles the global structure, while pixel-space super-resolution provides high-frequency details. This division of labor avoids the need to repeatedly perform a large number of denoising steps on a full-resolution grid, resulting in an overall design that is both simple and practically valuable.

Critical weaknesses
  • Dependence on external super-resolution models: The effectiveness of this method depends, to some extent, on the performance of the super-resolution network employed (e.g., Real-ESRGAN). If the SR model introduces significant artifacts, the subsequent single-step high-resolution refinement may not be able to fully recover the image.
Questions for authors
  • On FLUX.1-dev, MrFlow and SPEED exhibit different DPG-Bench performance at comparable speedup levels. Could authors further discuss the possible reasons for this difference?
4 comments
AuthorsOpenReview ↗

Response to Reviewer x744

We are deeply grateful for your support of our work, and we provide detailed responses to your comments as follows:

W1. Dependence on external super-resolution models: The effectiveness of this method depends, to some extent, on the performance of the super-resolution network employed (e.g., Real-ESRGAN). If the SR model introduces significant artifacts, the subsequent single-step high-resolution refinement may not be able to fully recover the image.

A1. Thank you for the feedback. The additional distortions introduced by the super-resolution network that you mention are precisely one of the focal points of MrFlow's design. This is consistent with our argument in Section 3.2 of the main text that GAN-based Real-ESRGAN should be used.

  • Although strategies that do not introduce additional artifacts also exist, such as direct linear interpolation, the high-frequency signals they lack are difficult to supplement directly through the subsequent noise-injection-and-denoising process. While GAN-based Real-ESRGAN supplements a large amount of high-frequency information and may also introduce additional artifacts, we point out in Sections 3.3 and 3.4 and Appendix Sections C.3 and C.5 that these artifacts are high-frequency details, which are precisely what a diffusion model is good at removing at the end of inference, closer to clean data, demonstrating the rationale behind the closely connected design choices in MrFlow.
  • An illustrative example is shown in Figure 3 of the original manuscript: when the Low Resolution image is fed into Real-ESRGAN and directly output after super-resolution, the character “m” instead appears incomplete; subsequent single-step high-resolution refinement effectively removes the introduced high-frequency defects. Figures 3 and 4 both visually demonstrate that the few artifacts remaining after super-resolution can be successfully and relatively consistently repaired by the final single high-resolution step, validating the stability of MrFlow in operation. The additional analysis is provided in Section 4.3 and Appendix Section C.2 of the revised manuscript.
AuthorsOpenReview ↗

Response to Reviewer x744

Q1. On FLUX.1-dev, MrFlow and SPEED exhibit different DPG-Bench performance at comparable speedup levels. Could authors further discuss the possible reasons for this difference?

A2. Thank you for the suggestion. Although MrFlow has higher speedup and generation quality in the vast majority of scenarios, as well as generally better overall visual quality as shown in Figures 1 and 7 of the original manuscript, its lower DPG-Bench score than SPEED on FLUX.1-dev deserves a more detailed explanation. Because diffusion generation is diverse, and SPEED and MrFlow in particular perform inference from initial noise at different resolutions, the difficulty of evaluation and the preferences of the scoring system affect the resulting metric values. Our specific explanation is as follows.

We first review the evaluation mechanism: DPG-Bench decomposes each prompt into questions about entities, attributes, and relationships, and organizes them as a dependency graph. The VQA model answers each question independently, assigning a score of 1 to an exact “yes” and 0 to any other answer. The evaluator then applies a dependency-gating mechanism: if a parent node receives a zero score, its child questions are also assigned zero, even if the child's own answer is “yes”; these zeros can propagate to deeper child nodes. The image score is the average score over all questions. Therefore, an uncertain judgment about a root concept may suppress the scores for many correctly recognized visual details.

We have added 2 sets of concrete examples in Appendix Section H.1 (Figure 9) of the revised manuscript to make this effect directly visible. For the politician prompt, MrFlow receives 15 affirmative answers out of 16, including those about the red jersey, volleyball, and podium. However, “Is there a politician?” receives “no”. This single root judgment suppresses 14 otherwise positive answers, leaving only the independent stage question: MrFlow’s accuracy of 15/16 is given a score of only 1/16 (6.25), whereas SPEED obtains 13/16 (81.25). The root question asks for an occupation that is difficult to establish from appearance alone. For the “Blind Ambition” prompt, the requested attribute is “their gaze obscured”: MrFlow visibly covers the eyes, whereas SPEED leaves them visible, yet the VQA answers to “Is the character’s gaze obscured?” are “no” and “yes”, respectively. In terms of human visual perception, I believe that MrFlow's generation quality is clearly better in at least the first set of examples, without the roughness and blurriness of SPEED when enlarged, yet it is given a lower score by the VQA model used in DPG-Bench.

The examples on Qwen-Image also reveal flaws in DPG-Bench evaluation. As shown in Figure 9 in Appendix Section H.1 of the revised manuscript, For the politician prompt, MrFlow generates a correct image with clear semantics and a clean appearance, and DPG-Bench also gives it a full score; however, although the image generated by SPEED is completely malformed, DPG-Bench still gives it a score of 15/16 (93.75).

Overall, although DPG-Bench is one of the metrics most widely used by our peers, assessing the generation quality of different acceleration strategies by considering other metrics together with actual generated examples should be a more comprehensive choice, and MrFlow consistently has better visual quality in actual generation.

Thank you for your detailed response and the additional experiments. The evaluation and analysis regarding DPG-Bench are thorough and have addressed my concerns in this area. While the introduction of an additional GAN ​​inevitably brings some negative side effects, the detailed supplementary analysis demonstrates that most of the primary drawbacks can be mitigated in subsequent stages. I will maintain my current score for now.

AuthorsOpenReview ↗

Response to Reviewer x744

We sincerely appreciate your careful review and your recognition of MrFlow’s practical value. We’re glad our additional analysis addressed your concerns.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely appreciate the positive feedback from the ACs and reviewers regarding MrFlow. To provide a clearer understanding of our work, we summarize our main contributions and the strengths acknowledged by the reviewers as follows:

We propose a simple and practical training-free strategy called MrFlow for accelerating generation with diffusion models. As every reviewer (x744, UhWG, gSxT) explicitly acknowledged in the first point of their Strengths, MrFlow can improve the generation efficiency of image diffusion models in a plug-and-play manner without training. From low-resolution generation to super-resolution upsampling in pixel space and then single-step high-resolution refinement after noise injection, the design of each stage of MrFlow has a clear and complete motivation, and the pipeline is clear and concise; ultimately, it achieves a better quality–speed trade-off than other training-free strategies on models such as Qwen-Image.

We have noted several concerns that reviewers have highlighted as particularly important:

  • Differences from typical multi-resolution strategies and noise-injection-based repainting strategies: As stated in the related work (lines 144 and 155) and the background section of the appendix (lines 725 and 755) of our original manuscript, MrFlow differs clearly from existing strategies in application domain, underlying motivation, and pipeline design; among similar pipelines, only MrFlow targets inference acceleration and provides the corresponding analysis of principles and design, with its design principle centering on assigning the unavoidable multi-step diffusion process to the low-resolution stage as much as possible, so that only a single high-resolution step is needed to complete the final detail refinement. The specific differences are as follows:
    • Multi-resolution strategies such as Hires.fix, HiFlow, and I-Max aim to extend and improve the generation capabilities of models, namely, to generate images beyond the recommended resolution natively supported by the models for inference. They allocate similar numbers of steps to their stages, take a long time to generate images, and use a noise strength greater than 0.5 at the transition. MrFlow, in contrast, focuses entirely on generation acceleration, and the final generated images are at the model's recommended resolution, without generating beyond the resolution supported by the original model. MrFlow particularly emphasizes that the final high-resolution stage requires only a single step, with a re-noising strength of around 0.15 at the transition.
    • Noise-injection-based repainting strategies such as SDEdit typically perform style transfer or detail editing tasks. Based on SDE-related theory, they receive complete images and repaint them with a noise strength typically greater than 0.5, using many timesteps and taking a long time to generate images. In MrFlow, adding noise and then denoising is intended to correct the erroneous high-frequency details left by the preceding fast low-resolution sampling and upsampling; we model and analyze this process using theory related to high- and low-frequency overwriting and identify a suitable noise strength of around 0.15, after which only a single denoising step is needed, making it very fast.
    • Based on the reviewers' suggestions, we have added results for strategies such as Hires.fix when generating at normal resolutions, as well as results obtained by replacing our direct small-amplitude noise injection and single-step denoising with SDEdit. The experiments show that MrFlow achieves clearly better generation quality and speed, demonstrating its distinctive design as an acceleration strategy.
  • Performance concerns and comparison fairness after applying a super-resolution model: Introducing the pretrained GAN-based Real-ESRGAN is one of the key design choices of MrFlow; it requires no additional continued training or modification, and its additional inference overhead is extremely low (0.6%). MrFlow observes that it effectively supplements the high-frequency signals missing from low-resolution generation, and finds that the small number of high-frequency errors it introduces are precisely what a diffusion model is good at removing at the end of inference, closer to clean data, thereby bringing significant gains in generation quality. Based on the reviewers' suggestions, we have added metrics such as MUSIQ and DeQA that better reveal the differences among super-resolution strategies, and have also reported the results of RALU and SPEED after switching to pixel-space upsampling with Real-ESRGAN. The experiments show that MrFlow is best suited to Real-ESRGAN, further corroborating the advantages of its upsampling strategy.

For a detailed explanation, please refer to our responses to each reviewer. We have also incorporated the additional content into the revised manuscript, with updates marked in red.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.