Noise-Started One-Step Real-World Super-Resolution via LR-Conditioned SplitMeanFlow and GAN Refinement

Wei Zhu¹     Kai Zhang^2,*     Yu Zheng¹     Lei Luo¹     Yong Guo³     Jian Yang^1,2,*

¹Nanjing University of Science and Technology     ²Nanjing University     ³Huawei

⏰ Update

• 2026.3.8: Create this repo.

⭐ If SCMSR is helpful to you, please help star this repo. Thanks!

🌟 Overview Framework

😍 Visual Results

⚙ Dependencies and Installation

## git clone this repository
git clone https://github.com/wzhu121/SMFSR.git
cd SMFSR

# create an environment with python >= 3.10
conda create -n SMFSR python=3.10
conda activate SMFSR
pip install -r requirements.txt 

🍭 Inference with script

Step 1: Download Checkpoints

• Download the [smfsr_f and smfsr_q] checkpoints and place them in the following directories: preset/smfsr_f and preset/smfsr_q.
• Download the [stable-diffusion-3.5-medium] checkpoints and place it in the preset/stable-diffusion-3.5-medium directory.
• Download the [clip-vit-large-patch14-336] and [llava-v1.5-13b] and place them in the llava_ckpt directory.

Step 2: Prepare testing data

You can download RealSR, DrealSR from [SeeSR], and download RealLQ250 from [DreamClear].

Step 3: Running testing command

# test w/o llava, one GPU is enough
bash scripts/test_wollava.sh

# test w/ llava, two GPUs are required
bash scripts/test_wllava.sh

🔥 Training

To be updated.

License

This project is released under the Apache 2.0 license.

Acknowledgement

This project is based on DiT4SR. Thanks for the awesome work!