September 8, 2019

928 words 5 mins read



ECCV18 Workshops - Enhanced SRGAN. Champion PIRM Challenge on Perceptual Super-Resolution (Third Region)

repo name xinntao/ESRGAN
repo link
language Python
size (curr.) 25454 kB
stars (curr.) 2310
created 2018-08-31
license Apache License 2.0

ESRGAN (Enhanced SRGAN) [BasicSR] [EDVR] [DNI]

We have merged the training codes of ESRGAN into MMSR :smile:

MMSR is an open source image and video super-resolution toolbox based on PyTorch. It is a part of the open-mmlab project developed by Multimedia Laboratory, CUHK. MMSR is based on our previous projects: BasicSR, ESRGAN, and EDVR.

We have simplified the network structure file. You can convert the previously save models (*.pth) with the script; If you want to use the old arch, you can find it here.

Check out our new work on:

  1. Video Super-Resolution: EDVR: Video Restoration with Enhanced Deformable Convolutional Networks, which has won all four tracks in NTIRE 2019 Challenges on Video Restoration and Enhancement (CVPR19 Workshops).
  2. DNI (CVPR19): Deep Network Interpolation for Continuous Imagery Effect Transition

Enhanced Super-Resolution Generative Adversarial Networks

By Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, Chen Change Loy

This repo only provides simple testing codes, pretrained models and the network strategy demo. For full training and testing codes, please refer to BasicSR.

We won the first place in PIRM2018-SR competition (region 3) and got the best perceptual index. The paper is accepted to ECCV2018 PIRM Workshop.

:triangular_flag_on_post: Add Frequently Asked Questions.

For instance,

  1. How to reproduce your results in the PIRM18-SR Challenge (with low perceptual index)?
  2. How do you get the perceptual index in your ESRGAN paper?


    author = {Wang, Xintao and Yu, Ke and Wu, Shixiang and Gu, Jinjin and Liu, Yihao and Dong, Chao and Qiao, Yu and Loy, Chen Change},
    title = {ESRGAN: Enhanced super-resolution generative adversarial networks},
    booktitle = {The European Conference on Computer Vision Workshops (ECCVW)},
    month = {September},
    year = {2018}

The RRDB_PSNR PSNR_oriented model trained with DF2K dataset (a merged dataset with DIV2K and Flickr2K (proposed in EDSR)) is also able to achive high PSNR performance.

Method Training dataset Set5 Set14 BSD100 Urban100 Manga109
SRCNN 291 30.48/0.8628 27.50/0.7513 26.90/0.7101 24.52/0.7221 27.58/0.8555
EDSR DIV2K 32.46/0.8968 28.80/0.7876 27.71/0.7420 26.64/0.8033 31.02/0.9148
RCAN DIV2K 32.63/0.9002 28.87/0.7889 27.77/0.7436 26.82/ 0.8087 31.22/ 0.9173
RRDB(ours) DF2K 32.73/0.9011 28.99/0.7917 27.85/0.7455 27.03/0.8153 31.66/0.9196

Quick Test


  • Python 3
  • PyTorch >= 1.0 (CUDA version >= 7.5 if installing with CUDA. More details)
  • Python packages: pip install numpy opencv-python

Test models

  1. Clone this github repo.
git clone
  1. Place your own low-resolution images in ./LR folder. (There are two sample images - baboon and comic).
  2. Download pretrained models from Google Drive or Baidu Drive. Place the models in ./models. We provide two models with high perceptual quality and high PSNR performance (see model list).
  3. Run test. We provide ESRGAN model and RRDB_PSNR model and you can config in the
  1. The results are in ./results folder.

Network interpolation demo

You can interpolate the RRDB_ESRGAN and RRDB_PSNR models with alpha in [0, 1].

  1. Run python 0.8, where 0.8 is the interpolation parameter and you can change it to any value in [0,1].
  2. Run python models/interp_08.pth, where models/interp_08.pth is the model path.

Perceptual-driven SR Results

You can download all the resutls from Google Drive. (:heavy_check_mark: included; :heavy_minus_sign: not included; :o: TODO)

HR images can be downloaed from BasicSR-Datasets.

Datasets LR ESRGAN SRGAN EnhanceNet CX
Set5 :heavy_check_mark: :heavy_check_mark: :heavy_check_mark: :heavy_check_mark: :o:
Set14 :heavy_check_mark: :heavy_check_mark: :heavy_check_mark: :heavy_check_mark: :o:
BSDS100 :heavy_check_mark: :heavy_check_mark: :heavy_check_mark: :heavy_check_mark: :o:
PIRM (val, test) :heavy_check_mark: :heavy_check_mark: :heavy_minus_sign: :heavy_check_mark: :heavy_check_mark:
OST300 :heavy_check_mark: :heavy_check_mark: :heavy_minus_sign: :heavy_check_mark: :o:
urban100 :heavy_check_mark: :heavy_check_mark: :heavy_minus_sign: :heavy_check_mark: :o:
DIV2K (val, test) :heavy_check_mark: :heavy_check_mark: :heavy_minus_sign: :heavy_check_mark: :o:


We improve the SRGAN from three aspects:

  1. adopt a deeper model using Residual-in-Residual Dense Block (RRDB) without batch normalization layers.
  2. employ Relativistic average GAN instead of the vanilla GAN.
  3. improve the perceptual loss by using the features before activation.

In contrast to SRGAN, which claimed that deeper models are increasingly difficult to train, our deeper ESRGAN model shows its superior performance with easy training.

Network Interpolation

We propose the network interpolation strategy to balance the visual quality and PSNR.

We show the smooth animation with the interpolation parameters changing from 0 to 1. Interestingly, it is observed that the network interpolation strategy provides a smooth control of the RRDB_PSNR model and the fine-tuned ESRGAN model.

Qualitative Results

PSNR (evaluated on the Y channel) and the perceptual index used in the PIRM-SR challenge are also provided for reference.

Ablation Study

Overall visual comparisons for showing the effects of each component in ESRGAN. Each column represents a model with its configurations in the top. The red sign indicates the main improvement compared with the previous model.

BN artifacts

We empirically observe that BN layers tend to bring artifacts. These artifacts, namely BN artifacts, occasionally appear among iterations and different settings, violating the needs for a stable performance over training. We find that the network depth, BN position, training dataset and training loss have impact on the occurrence of BN artifacts.

Useful techniques to train a very deep network

We find that residual scaling and smaller initialization can help to train a very deep network. More details are in the Supplementary File attached in our paper.

The influence of training patch size

We observe that training a deeper network benefits from a larger patch size. Moreover, the deeper model achieves more improvement (∼0.12dB) than the shallower one (∼0.04dB) since larger model capacity is capable of taking full advantage of larger training patch size. (Evaluated on Set5 dataset with RGB channels.)

comments powered by Disqus