Official PyTorch code for Hierarchical Conditional Flow: A Unified Framework for Image Super-Resolution and Image Rescaling (HCFlow, ICCV2021)

Last update: Dec 30, 2022

Overview

Hierarchical Conditional Flow: A Unified Framework for Image Super-Resolution and Image Rescaling (HCFlow, ICCV2021)

This repository is the official PyTorch implementation of Hierarchical Conditional Flow: A Unified Framework for Image Super-Resolution and Image Rescaling (arxiv, supp).

🚀 🚀 🚀 News:

Aug. 17, 2021: See our recent work for spatially variant kernel estimation: Mutual Affine Network for Spatially Variant Kernel Estimation in Blind Image Super-Resolution (MANet), ICCV2021
Aug. 17, 2021: See our recent work for real-world image SR: Designing a Practical Degradation Model for Deep Blind Image Super-Resolution (BSRGAN), ICCV2021
Aug. 17, 2021: See our previous flow-based work: Flow-based Kernel Prior with Application to Blind Super-Resolution (FKP), CVPR2021.

Normalizing flows have recently demonstrated promising results for low-level vision tasks. For image super-resolution (SR), it learns to predict diverse photo-realistic high-resolution (HR) images from the low-resolution (LR) image rather than learning a deterministic mapping. For image rescaling, it achieves high accuracy by jointly modelling the downscaling and upscaling processes. While existing approaches employ specialized techniques for these two tasks, we set out to unify them in a single formulation. In this paper, we propose the hierarchical conditional flow (HCFlow) as a unified framework for image SR and image rescaling. More specifically, HCFlow learns a bijective mapping between HR and LR image pairs by modelling the distribution of the LR image and the rest high-frequency component simultaneously. In particular, the high-frequency component is conditional on the LR image in a hierarchical manner. To further enhance the performance, other losses such as perceptual loss and GAN loss are combined with the commonly used negative log-likelihood loss in training. Extensive experiments on general image SR, face image SR and image rescaling have demonstrated that the proposed HCFlow achieves state-of-the-art performance in terms of both quantitative metrics and visual quality.

Requirements

Python 3.7, PyTorch == 1.7.1
Requirements: opencv-python, lpips, natsort, etc.
Platforms: Ubuntu 16.04, cuda-11.0

cd HCFlow-master
pip install -r requirements.txt

Quick Run (takes 1 Minute)

To run the code with one command (without preparing data), run this command:

cd codes
# face image SR
python test_HCFLow.py --opt options/test/test_SR_CelebA_8X_HCFlow.yml

# general image SR
python test_HCFLow.py --opt options/test/test_SR_DF2K_4X_HCFlow.yml

# image rescaling
python test_HCFLow.py --opt options/test/test_Rescaling_DF2K_4X_HCFlow.yml

Data Preparation

The framework of this project is based on MMSR and SRFlow. To prepare data, put training and testing sets in ./datasets as ./datasets/DIV2K/HR/0801.png. Commonly used SR datasets can be downloaded here. There are two ways for accerleration in data loading: First, one can use ./scripts/png2npy.py to generate .npy files and use data/GTLQnpy_dataset.py. Second, one can use .pklv4 dataset (recommended) and use data/LRHR_PKL_dataset.py. Please refer to SRFlow for more details. Prepared datasets can be downloaded here.

Training

To train HCFlow for general image SR/ face image SR/ image rescaling, run this command:

cd codes

# face image SR
python train_HCFLow.py --opt options/train/train_SR_CelebA_8X_HCFlow.yml

# general image SR
python train_HCFLow.py --opt options/train/train_SR_DF2K_4X_HCFlow.yml

# image rescaling
python train_HCFLow.py --opt options/train/train_Rescaling_DF2K_4X_HCFlow.yml

All trained models can be downloaded from here.

Testing

Please follow the Quick Run section. Just modify the dataset path in test_HCFlow_*.yml.

Results

We achieved state-of-the-art performance on general image SR, face image SR and image rescaling.

For more results, please refer to the paper and supp for details.

Citation

@inproceedings{liang21hcflow,
  title={Hierarchical Conditional Flow: A Unified Framework for Image Super-Resolution and Image Rescaling},
  author={Liang, Jingyun and Lugmayr, Andreas and Zhang, Kai and Danelljan, Martin and Van Gool, Luc and Timofte, Radu},
  booktitle={IEEE Conference on International Conference on Computer Vision},
  year={2021}
}

License & Acknowledgement

This project is released under the Apache 2.0 license. The codes are based on MMSR, SRFlow, IRN and Glow-pytorch. Please also follow their licenses. Thanks for their great works.

Comments

Testing without GT

Is there a way to run the test without GT? I just want to infer the model. I found a mode called LQ which -I think- should only load the images in LR directory. But this mode gives me the error: assert real_crop * self.opt['scale'] * 2 > self.opt['kernel_size'] TypeError: '>' not supported between instances of 'int' and 'NoneType'

in LQ_dataset.py", line 88
solved ✅

opened by AhmedHashish123 4
Add Docker environment & web demo

Hey @JingyunLiang !👋

This pull request makes it possible to run your model inside a Docker environment, which makes it easier for other people to run it. We're using an open source tool called Cog to make this process easier.

This also means we can make a web page where other people can try out your model! View it here: https://replicate.ai/jingyunliang/hcflow-sr, which currently supports Image Super-Resolution.

Claim your page here so you can edit it, and we'll feature it on our website and tweet about it too.

In case you're wondering who I am, I'm from Replicate, where we're trying to make machine learning reproducible. We got frustrated that we couldn't run all the really interesting ML work being done. So, we're going round implementing models we like. 😊

opened by chenxwh 2
The code implementation and the paper description seem different

Hi, your work is excellent, but there is one thing I don't understand.

What is written in the paper is:

"A diagonal covariance matrix with all diagonal elements close to zero"

But the code implementation in HCFlowNet_SR_arch.py line 64 is: basic. Gaussian diag.logp (LR, - torch. Ones_ like(lr)*6, fake_ lr_ from_ hr)

why use - torch. Ones_ like(lr)*6 as covariance matrix? This seems to be inconsistent with the description in the paper

opened by xmyhhh 2
environment

ImportError: /home/hbw/gcc-build-5.4.0/lib64/libstdc++.so.6: version `GLIBCXX_3.4.22' not found (required by /home/hbw/anaconda3/lib/python3.8/site-packages/scipy/fft/_pocketfft/pypocketfft.cpython-38-x86_64-linux-gnu.so)

Is this error due to my GCC version being too low, and your version is? looking forward to your reply!

opened by hbw945 2
Code versions of BRISQUE and NIQE used in paper

Hi, I have run performance tests with the Matlab versions of the NIQE and BRISQUE codes and found deviations from the values reported in the paper. Could you please provide a link to the code you used? thanks a lot~
solved ✅

opened by xmyhhh 1
Update on Replicate demo
Hello again @JingyunLiang :),

This pull request does a few little things:

Updated the demo link with an icon in README as you suggested

A bugfix for cleaning temporary directory on cog

We have added more functionality to the Example page of your model, now you can add and delete to customise the example gallery as you like (as the owner of the page)

Also, you could run cog push if you like to update the model of any other models on replicate in the future 😄
opened by chenxwh 1
About training and inference time?

Thanks for your nice work!

I want to know how much time do you need to train and inference with your models.

Furthermore, will information about params / FLOPs be reported?

Thanks.
solved ✅

opened by TiankaiHang 1
RuntimeError: The size of tensor a (20) must match the size of tensor b (40) at non-singleton dimension 3

Hi, I've encountered the error when I trained the HCFlowNet. I changed my ".png" dataset to ".pklv4" dataset. I was trained on the platform of windows 10 with 1 single GPU. Could you please help me find the error? Thanks a lot.

opened by William9Baker 0
How to build an invertible mapping between two variables whose dimensions are different ?

Maybe this is a stupid question, but I have been puzzled for quite a long time. In the image super-resolution task, the input and output have different dimensions. How to build an invertible mapping between them? I notice that you calculate the determinant of the Jacobian, so I thought the mapping here is strictly invertible?

opened by Wangbk-dl 0
How to make an invertible mapping between two variables whose dimensions are different ?

Maybe this is a stupid question, but I have been puzzled for quite a long time. In the image super-resolution task, the input and output have different dimensions. How to build such an invertible mapping between them ? Take an example: If I have a low-resolution(LR) image x, and I have had an invertible function G. I can feed LR image x into G, and generate an HR image y. But can you ensure that we could obtain an output the same as x when we feed y into G_inverse?

y = G(x) x' = G_inverse(y) =? x

I would appreciate it if you could offer some help.

opened by Wangbk-dl 0
New Super-Resolution Benchmarks
Hello,

MSU Graphics & Media Lab Video Group has recently launched two new Super-Resolution Benchmarks.

Video Upscalers Benchmark: Quality Enhancement determines the best upscaling methods for increasing video resolution and improving visual quality.

Super-Resolution for Video Compression benchmark aims to test Super-Resolution methods on compressed videos and select the best model for each video codec standard.

If you are interested in participating, you can add your algorithm following the submission steps:

Submit for Video Upscalers Benchmark: Quality Enhancement

Submit for Super-Resolution for Video Compression benchmark

We would be grateful for your feedback on our work!
opened by EvgeneyBogatyrev 0
Why NLL is negative during the training？

Great work! During the training process, we found that the output NLL is negative. But theoretically, NLL should be positive. Is there any explanation for this?

opened by IMSEMZPZ 0