PyTorch Implementation of the Model and Training Process of Real-ESRGAN

• DIV2K
• Flickr2K
• OutdoorSceneTraining (OST)
  I'll organize these three datasets, and after removing images with height or width less than 256, I'll compress them into a file called "train_HR.zip."

A weight file for a generator and discriminator model trained for 43500 iterations can be download from here and here(Google Drive Link).

python train.py
optional arguments:
  -h, --help            show this help message and exit
  --crop_size CROP_SIZE
                        training images crop size
  --upscale_factor UPSCALE_FACTOR
                        super resolution upscale factor
  --batch_size BATCH_SIZE
                        batch size of train dataset
  --warmup_batches WARMUP_BATCHES
                        number of batches with pixel-wise loss only
  --n_batches N_BATCHES
                        number of batches of training
  --residual_blocks RESIDUAL_BLOCKS
                        number of residual blocks in the generator
  --batch BATCH         batch to start training from
  --lr LR               adam: learning rate
  --sample_interval SAMPLE_INTERVAL
                        interval between saving image samples

• patch size : 256
• batch size : 16
• optimizer : Adam
• First train generator for 10k iteration with learning rate $2\times 10^{-4}$
• Then train generator and discriminator for 3.5k iteration with learning rate: 1e-4
• adopt exponential moving average(EMA) with beta = 0.999
• discriminator use Unet with SN(spectral normalization)
• Real Esrgan is trained with a combination loss of pixel-loss(L1 loss), perceptual loss and GAN loss
• use {conv1, ..., conv5} feature maps with weights {0.1, 0.1, 1, 1, 1} before activation in the pretrained VGG19 network as perceptual loss

python test_image.py  

optional arguments:
  -h, --help            show this help message and exit
  --upscale_factor UPSCALE_FACTOR
                        super resolution upscale factor
  --test_mode {CPU,GPU}
                        using CPU or GPU
  --image_name IMAGE_NAME
                        test low resolution image name
  --model_name MODEL_NAME
                        generator model epoch name

Upscale Factor = 4
low quality image:

REAL-ESRGAN results:

low quality image:

REAL-ESRGAN results:

low quality image:

REAL-ESRGAN results:

Upscale Factor = 4
This is a batch of images saved every 100 batches.The leftmost column is the low-resolution image obtained by interpolation using the BICUBIC method. The middle column is the original high-resolution image, and the rightmost column is the super-resolution image reconstructed using the model.