MFSR: MeanFlow Distillation for One-Step
Real-World Image Super-Resolution

Ruiqing Wang¹   Kai Zhang^1†   Yuanzhi Zhu²   Hanshu Yan³   Shilin Lu⁴   Jian Yang¹

¹ Nanjing University    ² LIX, École Polytechnique, CNRS, IPP
³ Salesforce AI Research    ⁴ Nanyang Technological University
^† Corresponding author

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 key insight is to use MeanFlow as the learning target, enabling the student to approximate the average velocity between arbitrary states and thus capture teacher dynamics without requiring explicit rollouts. We further improve upon standard Classifier-Free Guidance by incorporating teacher CFG predictions as stronger supervisory signals. MFSR achieves perceptual quality on par with or even better than multi-step teacher models, while dramatically reducing inference cost.

Motivation

Diffusion- and flow-based models achieve impressive restoration quality for Real-ISR, but require many sampling steps that make deployment impractical. Existing one-step distillation methods suffer from a fundamental trade-off: they reduce inference cost but degrade perceptual quality and sacrifice the ability to run additional refinement steps when needed. We ask: can we distill a flow-based SR model into a one-step student that matches teacher quality and still supports multi-step refinement?

Method

MFSR distills a pretrained DiT4SR flow teacher into a student model using MeanFlow as the learning objective. Rather than matching instantaneous velocities (as in consistency distillation), the student learns to predict the average velocity over any time interval [t, s]. This single unified target enables both one-step inference and optional multi-step refinement without any architectural changes.

To strengthen supervision, we introduce Teacher CFG Distillation: instead of using naive CFG outputs, we incorporate the teacher's full classifier-free guidance predictions as supervisory signals, preserving fine detail and high-frequency texture that vanilla distillation tends to lose.

BibTeX

If you find this work helpful, please cite:

@article{wang2026mfsr,
  title={MFSR: MeanFlow Distillation for One Step Real-World Image Super Resolution},
  author={Wang, Ruiqing and Zhang, Kai and Zhu, Yuanzhi and Yan, Hanshu and Lu, Shilin and Yang, Jian},
  journal={arXiv preprint arXiv:2603.20690},
  year={2026}
}
    

Project page template is borrowed from HiFlow.