Multi-Scale Deep Super-Resolution System (MDSR)

MDSR model pre-trained on DIV2K (800 images training, augmented to 4000 images, 100 images validation) for 2x, 3x and 4x image super resolution. It was introduced in the paper Enhanced Deep Residual Networks for Single Image Super-Resolution by Lim et al. (2017) and first released in this repository.

The goal of image super resolution is to restore a high resolution (HR) image from a single low resolution (LR) image. The image below shows the ground truth (HR), the bicubic upscaling and model upscaling.

Comparing Bicubic upscaling against the models x4 upscaling on Set5 Image 4

Model description

The MDSR is a model that uses both deeper and wider architecture (32 ResBlocks and 256 channels) to improve performance. It uses both global and local skip connections, and up-scaling is done at the end of the network. It doesn't use batch normalization layers (input and output have similar distributions, normalizing intermediate features may not be desirable) instead it uses constant scaling layers to ensure stable training. An L1 loss function (absolute error) is used instead of L2 (MSE), the authors showed better performance empirically and it requires less computation.

This model also applies the balanced attention (BAM) method invented by Wang et al. (2021) to further improve the results.

Intended uses & limitations

You can use the pre-trained models for upscaling your images 2x, 3x and 4x. You can also use the trainer to train a model on your own dataset.

How to use

The model can be used with the super_image library:

pip install super-image

Here is how to use a pre-trained model to upscale your image:

from super_image import MdsrModel, ImageLoader
from PIL import Image
import requests

url = ''
image =, stream=True).raw)

model = MdsrModel.from_pretrained('eugenesiow/mdsr-bam', scale=2)      # scale 2, 3 and 4 models available
inputs = ImageLoader.load_image(image)
preds = model(inputs)

ImageLoader.save_image(preds, './scaled_2x.png')                        # save the output 2x scaled image to `./scaled_2x.png`
ImageLoader.save_compare(inputs, preds, './scaled_2x_compare.png')      # save an output comparing the super-image with a bicubic scaling

Open In Colab

Training data

The models for 2x, 3x and 4x image super resolution were pretrained on DIV2K, a dataset of 800 high-quality (2K resolution) images for training, augmented to 4000 images and uses a dev set of 100 validation images (images numbered 801 to 900).

Training procedure


We follow the pre-processing and training method of Wang et al.. Low Resolution (LR) images are created by using bicubic interpolation as the resizing method to reduce the size of the High Resolution (HR) images by x2, x3 and x4 times. During training, RGB patches with size of 64×64 from the LR input are used together with their corresponding HR patches. Data augmentation is applied to the training set in the pre-processing stage where five images are created from the four corners and center of the original image.

We need the huggingface datasets library to download the data:

pip install datasets

The following code gets the data and preprocesses/augments the data.

from datasets import load_dataset
from import EvalDataset, TrainDataset, augment_five_crop

augmented_dataset = load_dataset('eugenesiow/Div2k', 'bicubic_x4', split='train')\
    .map(augment_five_crop, batched=True, desc="Augmenting Dataset")                                # download and augment the data with the five_crop method
train_dataset = TrainDataset(augmented_dataset)                                                     # prepare the train dataset for loading PyTorch DataLoader
eval_dataset = EvalDataset(load_dataset('eugenesiow/Div2k', 'bicubic_x4', split='validation'))      # prepare the eval dataset for the PyTorch DataLoader


The model was trained on GPU. The training code is provided below:

from super_image import Trainer, TrainingArguments, MdsrModel, MdsrConfig

training_args = TrainingArguments(
    output_dir='./results',                 # output directory
    num_train_epochs=1000,                  # total number of training epochs

config = MdsrConfig(
    scale=4,                                # train a model to upscale 4x
    bam=True,                               # apply balanced attention to the network
model = MdsrModel(config)

trainer = Trainer(
    model=model,                         # the instantiated model to be trained
    args=training_args,                  # training arguments, defined above
    train_dataset=train_dataset,         # training dataset
    eval_dataset=eval_dataset            # evaluation dataset


Open In Colab

Evaluation results

The evaluation metrics include PSNR and SSIM.

Evaluation datasets include:

The results columns below are represented below as PSNR/SSIM. They are compared against a Bicubic baseline.

Dataset Scale Bicubic mdsr-bam
Set5 2x 33.64/0.9292 38/0.9607
Set5 3x 30.39/0.8678 35.07/0.9402
Set5 4x 28.42/0.8101 32.19/0.8949
Set14 2x 30.22/0.8683 33.68/0.9182
Set14 3x 27.53/0.7737 31.04/0.8582
Set14 4x 25.99/0.7023 28.73/0.7847
BSD100 2x 29.55/0.8425 33.77/0.9253
BSD100 3x 27.20/0.7382 29.62/0.8188
BSD100 4x 25.96/0.6672 28.5/0.7645
Urban100 2x 26.66/0.8408 32.04/0.9272
Urban100 3x 29.16/0.8717
Urban100 4x 23.14/0.6573 26.02/0.7834

Comparing Bicubic upscaling against the models x4 upscaling on Set5 Image 2

You can find a notebook to easily run evaluation on pretrained models below:

Open In Colab

BibTeX entry and citation info

    title={BAM: A Lightweight and Efficient Balanced Attention Mechanism for Single Image Super Resolution}, 
    author={Fanyi Wang and Haotian Hu and Cheng Shen},
  title={Fast, Accurate, and Lightweight Super-Resolution with Cascading Residual Network},
  author={Ahn, Namhyuk and Kang, Byungkon and Sohn, Kyung-Ah},
  journal={arXiv preprint arXiv:1803.08664},
Downloads last month
Hosted inference API

Unable to determine this model’s pipeline type. Check the docs .