- 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
-
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.
-
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.
-
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
-
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.
-
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.
-
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
- Could you provide quantitative and qualitative comparisons with Hires.fix/SDEdit-style baselines under the same step budget?
- Could you provide qualitative comparisons with RALU and SPEED using the interpolation-based MrFlow, without the external SR network?
- Could you report a quantitative sweep over σ_t and show whether a performance transition occurs around σ*?
10 comments
Response to Reviewer gSxT
Differences from noise-injection-based repainting strategies such as SDEdit:
- 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.
- 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.
- 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.
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:
- 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
- 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
- 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.
Response to Reviewer gSxT
Thank you for your careful review of our work. And we provide detailed responses to your comments as follows:
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: