---
title: RFMSR Super-Resolution
emoji: 🔍
colorFrom: red
colorTo: gray
sdk: gradio
sdk_version: 6.22.0
app_file: app.py
short_description: Residual Flow Matching for 4x image super-resolution
python_version: "3.12"
startup_duration_timeout: 30m
---

# RFMSR: Residual Flow Matching for Image Super-Resolution

This Space demos [RFMSR](https://huggingface.co/papers/2607.12753), a flow-matching-based image super-resolution model.

Upload a low-quality image and get a 4× upscaled result. The model performs residual flow matching in the latent space of the SD2.1 VAE, guided by DINOv2 semantic features.

## Usage

1. Upload a low-quality image (or click an example)
2. Adjust the upscale factor and number of steps if desired
3. Click "Super-Resolve"

**Multi-step mode** (default, 15 steps) produces sharp, high-quality results. **One-step mode** (1 step) is faster but produces softer/blurrier output.

## Links

- [Paper (arXiv 2607.12753)](https://arxiv.org/abs/2607.12753)
- [GitHub](https://github.com/Faze-Hsw/RFMSR)
- [Hugging Face Model](https://huggingface.co/frozen2001/RFMSR)