MFSR: MeanFlow Distillation for One Step Real-World Image Super Resolution
Ruiqing Wang, Kai Zhang, Yuanzhi Zhu, Hanshu Yan, Shilin Lu, Jian Yang
TL;DR
MFSR introduces a one-step distillation method for real-world image super-resolution that maintains high quality and allows optional refinement, significantly reducing computational cost compared to multi-step models.
Contribution
The paper proposes MeanFlow Distillation (MFSR), a novel framework enabling single-step super-resolution with optional multi-step refinement, improving efficiency and quality over existing multi-step diffusion models.
Findings
MFSR achieves comparable or better results than multi-step models.
The method significantly reduces inference time and computational cost.
Enhanced preservation of fine details through teacher CFG distillation.
Abstract
Diffusion- and flow-based models have advanced Real-world Image Super-Resolution (Real-ISR), but their multi-step sampling makes inference slow and hard to deploy. One-step distillation alleviates the cost, yet often degrades restoration quality and removes the option to refine with more steps. We present Mean Flows for Super-Resolution (MFSR), a new distillation framework that produces photorealistic results in a single step while still allowing an optional few-step path for further improvement. Our approach uses MeanFlow as the learning target, enabling the student to approximate the average velocity between arbitrary states of the Probability Flow ODE (PF-ODE) and effectively capture the teacher's dynamics without explicit rollouts. To better leverage pretrained generative priors, we additionally improve original MeanFlow's Classifier-Free Guidance (CFG) formulation with teacher CFG…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1. Well-motivated problem: The paper addresses a practical challenge in real-world super-resolution - reducing inference cost while maintaining quality. The motivation for adapting MeanFlow to SR distillation is clear and reasonable. 2. Flexible inference scheme: Unlike existing one-step SR methods, MFSR preserves the ability to perform few-step sampling, providing a controllable trade-off between speed and quality. This is a valuable feature for practical deployment. 3. Simplified training pipe
1. Inconsistent and concerning quantitative results: (1) MFSR shows significantly lower PSNR/SSIM across all benchmarks (e.g., 21.25 vs 24.50 PSNR on DIV2K-Val for SinSR). While the authors mention perception-distortion tradeoff, the gap is substantial. (2) The paper dismisses these metrics as "misaligned with human perception" but still reports them extensively, creating confusion about which metrics to trust. (3) On some metrics (LPIPS, DISTS), MFSR performs worse than several baselines, cont
1. The proposed method, MFSR, reduces computational cost while maintaining quality. The use of the MeanFlow distillation strategy enables high-quality one-step restoration and retains flexibility for optional few-step refinement. 2. The idea of leveraging a pre-trained teacher model to guide the student model in distilling knowledge is well-executed. This reduces convergence time and improves the quality of restoration, making the approach more practical for real-world applications. 3. The mod
1. The main difference between MFSR and MeanFlow lies in the use of teacher CFG to improve instantaneous velocity. Therefore, the technical contribution appears incremental rather than fundamentally novel. 2. The quantitative results reported show limited competitiveness compared with other methods. MFSR fails to achieve state-of-the-art performance on all FR metrics. While I understand that PSNR and SSIM are less meaningful in Real-ISR, LPIPS and DISTS remain important reference indicators. Am
1. The authors propose a new view on Mean Flow, not as a training from scratch approach, but as a distillation approach using the instantaneous velocity predicted by the teacher instead of a one-sample estimator. 2. Good quality of the resulting model.
1. **Limited novelty and lack of SR-specific contribution.** While the method differs from the previously proposed Mean Flow approach by using instantaneous velocity provided by the teacher model, it introduces no super-resolution-specific modeling or insight. As a result, it does not meaningfully address the distinct challenges of the SR domain, and the contribution reads more as an engineering transfer rather than SR-oriented innovation. 2. **Insufficient analysis and insight.** The paper doe
(i) The authors reformulate classifier-free guidance for MeanFlow [4] to a distillation regime by leveraging a pretrained teacher during training. This CFG-based MeanFlow distillation yields stronger supervision than the original MeanFlow CFG, improving one and few step restoration quality. (ii) The approach is scaled to a DiT architecture [1,2,3], using a pretrained DiT4SR [3] both as the teacher and for student initialization. This scaling to relatively big models is nontrivial and demonstrat
(i) Using MeanFlow with distillation and CFG seems like a natural next step. While useful, this idea is not very new or surprising. The paper feels more like an engineering improvement than a theoretical one. (ii) The authors say metrics don’t reflect real-world SR well and use a user study to support this. However, the human evaluation compares only a few methods, so the results are not very strong. A larger study with more methods would help. Since the paper focuses on practical SR, it should
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Image Processing Techniques · Advanced Vision and Imaging · Generative Adversarial Networks and Image Synthesis