ResFlow-Unofficial

Unofficial PyTorch implementation starter for Reversing Flow for Image Restoration (CVPR 2025).

If this repo saves you reading / reproduction time, please give it a star and follow @StaryMoon. It helps me keep building open reproduction starters for recent restoration papers.

Status

This repository is an independent, unofficial, work-in-progress reproduction starter.

• Paper: Reversing Flow for Image Restoration
• Venue: CVPR 2025, pp. 7545-7558
• Official code: the CVPR page says the authors plan to release code; this repository is not affiliated with the authors.
• Reproduction status: benchmarks are not reproduced yet.

The first release focuses on making the core idea easy to inspect:

• conditional velocity network for image restoration
• degradation operators for toy experiments
• flow-matching style training objective
• few-step Euler sampling loop from degraded image to restored image
• smoke-test script for quick sanity checks

What This Is

ResFlow models image restoration as a deterministic path from low-quality images to high-quality images and learns a velocity field along that path. This repo implements a compact starter version of that idea:

degraded image x_0  -- learned velocity field -->  restored image x_1

The implementation is intentionally small so students and researchers can read it quickly, modify it, and plug in real datasets later.

What This Is Not

This is not the official codebase and not a drop-in reproduction of the paper's reported numbers.

Current limitations:

• exact architecture from the paper is not implemented yet
• auxiliary-variable details are simplified in v0.1.0
• no official checkpoints
• no full dataset pipeline
• no PSNR / SSIM benchmark table yet

Quick Start

git clone https://github.com/StaryMoon/ResFlow-Unofficial.git
cd ResFlow-Unofficial
pip install -r requirements.txt
python scripts/smoke_test.py

Expected output:

loss: ...
restored: torch.Size([2, 3, 64, 64])

Minimal Usage

import torch

from resflow_unofficial import RestorationFlow, random_degrade

clean = torch.rand(2, 3, 64, 64)
degraded = random_degrade(clean)

model = RestorationFlow(image_channels=3, base_channels=32)
loss = model.training_loss(clean, degraded)
loss.backward()

restored = model.sample(degraded, steps=4)

Roadmap

• Match the exact ResFlow architecture more closely.
• Add auxiliary-state modeling instead of the simplified conditional path.
• Add dataset loaders for deraining, deblurring, denoising, and low-light enhancement.
• Add PSNR / SSIM evaluation scripts.
• Reproduce a small benchmark table.
• Add pretrained checkpoints when experiments become stable.

Why I Built This

Image restoration papers often become useful only after someone turns the math into a runnable repo. This project is meant to be that first public scaffold: clear enough to read, small enough to fork, and honest about what is still missing.

Citation

If you use the method, please cite the original paper:

@InProceedings{Qin_2025_CVPR,
  author = {Qin, Haina and Luo, Wenyang and Wang, Libin and Zheng, Dandan and Chen, Jingdong and Yang, Ming and Li, Bing and Hu, Weiming},
  title = {Reversing Flow for Image Restoration},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2025},
  pages = {7545--7558}
}

License

This repository is released under the MIT License. The paper and any official materials remain owned by their respective authors / publishers.