RFMSR: Residual Flow Matching for Image Super-Resolution

This repository contains the weights for RFMSR (Residual Flow Matching for Image Super-Resolution), a vision-only super-resolution framework that utilizes a residual flow design centering the source distribution at the low-quality (LQ) latent to preserve structural priors.

📚 Paper | 💻 GitHub Repository

Method Overview

RFMSR centers the source distribution at the LR latent, reducing the transport distance and preserving structural priors throughout the flow trajectory. It uses a two-phase training strategy to achieve high-quality single-step generation without sacrificing multi-step refinement.

Pretrained Weights

This repository hosts the following model checkpoints:

• rfmsr.safetensors — Phase I: Multi-step Flow Matching model (15-step recommended)
• rfmsr_os.safetensors — Phase II: One-step model
• rfmsr_consistency.safetensors — Consistency distillation: One-step model

Quick Start (Inference)

To run inference, please clone the official GitHub repository and install the dependencies:

git clone https://github.com/Faze-Hsw/RFMSR.git
cd RFMSR
conda create -n rfmsr python=3.12 -y
conda activate rfmsr
pip install -r requirements.txt

Download the checkpoints via huggingface-cli:

pip install huggingface_hub
huggingface-cli download frozen2001/RFMSR ckpts/ --local-dir . --local-dir-use-symlinks False

Run super-resolution inference on your images:

# One-step inference (fast, recommended)
python infer_rfmsr.py --input input.png --steps 1

# Multi-step inference (higher quality)
python infer_rfmsr.py --input input.png --steps 15

Citation

If you find this work useful, please cite:

@misc{huang2026rfmsrresidualflowmatching,
      title={RFMSR: Residual Flow Matching for Image Super-Resolution},
      author={Shuwei Huang and Tianyao Luo and Jicheng Liu and Daizong Liu and Pan Zhou},
      year={2026},
      eprint={2607.12753},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.12753},
}