Following are the original trained models that authors made available for the different architectures, compiled here only for convenience. Also some experimental models that have been trained using this repository for use as pretrained models have been added.

Additional custom models can be found in the model database.

In order to use these models, they can be downloaded and saved in any directory and the path should be added to the configuration file (usually pretrain_model_G). The default path to use is experiments/pretrained_models.

Super Resolution

Name	Models	Short Description	Source
ESRGAN	4x_RRDB_ESRGAN.pth, 4x_RRDB_ESRGAN_modarch.pth	the final ESRGAN model used in the paper and the modified architecture version from current repo	Google Drive
	4x_RRDB_PSNR.pth, 4x_RRDB_PSNR_modarch.pth	model with high PSNR performance and the modified architecture version from current repo
SRGAN*1	SRGAN_bicx4_303_505.pth	4x SRGAN model (with modification), trained on DIV2K, w/o BN, bicubic downsampling.	Google Drive
SRResNet*2	**multiple, see notes below**, SRResNet_bicx4_in3nf64nb16.pth	4x SRResNet model (with modification), trained on DIV2K, w/o BN, bicubic downsampling.	Google Drive
PPON	PPON.pth	PPON model presented in the paper	Google Drive
PAN	PAN.pth	4x pretrained modified PAN model with self-attention	Google Drive
SRFLOW	4x_srflow.pth	4x SRFlow model trained using a ESRGAN model from the original architecture as base (not fully trained)	Google Drive
	srflow_original.zip	the original SRFlow models at 4x and 8x scales (used the modified ESRGAN arch as base model)
SFTGAN	segmentation_OST_bic.pth	segmentation model for bicubic downsampled images, outdoor scenes	Google Drive
	sft_net_ini.pth	initialized SFTGAN model, initializing the SR generator with SRGAN_bicx4_303_505 parameters
	sft_net_torch.pth	torch version of SFTGAN model (paper)
	SFTGAN_bicx4_noBN_OST_bg.pth	PyTorch version of SFTGAN model: trained on OST dataset and use DIV2K as background images, w/o BN, bicubic downsampling

Image to image translation

These initial models are the same as the ones in the original pix2pix and CycleGAN repo and use the default UNet and ResNet networks configuration respectively (ie. using ConvTranspose2d for upsample layers).

Name	Models	Short Description	Source
pix2pix	facades_label2photo.pth	models originally available in the pix2pix repo. These correspond to the Pix2pix datasets.	Google Drive
	sat2map.pth
	map2sat.pth
	edges2shoes.pth
	edges2handbags.pth
	day2night.pth
CycleGAN	facades_label2photo.pth, facades_photo2label.pth	models originally available in the CycleGAN repo. These correspond to the CycleGAN datasets. In some cases, only one generator of the cycle was provided, but the missing generator can be trained with the original dataset.	Google Drive
	sat2map.pth, map2sat.pth
	horse2zebra.pth, zebra2horse.pth
	summer2winter_yosemite.pth, winter2summer_yosemite.pth
	cityscapes_photo2label.pth, cityscapes_label2photo.pth
	apple2orange.pth, orange2apple.pth
	monet2photo.pth, photo2monet.pth
	photo2ukiyoe.pth
	photo2cezanne.pth
	photo2vangogh.pth
	iphone2dslr_flower.pth
WBC	wbc.pth	'wbc.pth' is the current pretrained model trained using this code (new versions will be made available with further training). 'wbc_tf_converted.pth' model was converted from the original TensorFlow model in the WBC repo to PyTorch. To be able to interpolate models, other models used it as pretrained, but interpolations with the converted model won't work correctly because of differing padding and upscaling configurations.	Google Drive
	wbc_tf_converted.pth

Video (Experimental)

Following are some video models that have been trained using the networks available in this repository, many of them are experimental, but are useful to use as pretrained models for testing.

Name	Models	Short Description	Source
SOFVSR	SOFVSR.pth	4x pretrained SOFVSR model, using 3 frames	Google Drive
SOFVESRGAN	SOFVESRGAN.pth	4x pretrained modified SOFVSR model using ESRGAN network for super-resolution, using 3 frames	Google Drive
EVSRGAN	4x_EVSRGAN_REDS_pretrained.pth	4x EVSRGAN Pretrained using 3 frames and default arch options	pcloud
DVD	DVD_REDS-Deinterlace-*_G.pth	Real-time Deep Video Deinterlacing	pcloud
RIFE	RIFE.pth	Converted pretrained RIFE model from the three original pickle files into a single pth model	Google Drive

---

Below are some additional notes on some models:

SRResNet (EDSR)

Through experiments it was found that using:

• no batch normalization
• residual block style: Conv-ReLU-Conv

are the best network settings for this network.

Qualitative results [PSNR/dB]

Besides the mentioned SRResNet_bicx4_in3nf64nb16.pth, other pretrained SRResNet models can be downloaded from the Google Drive folder, with where trained using different network configurations. Below is a table that compares the results of each configuration (note the name convention, also explained after the table, to identify each configuration):

Model	Scale	Channel	DIV2K2	Set5	Set14	BSD100	Urban100
SRResNet_bicx2_in3nf64nb161	2	RGB	34.7203	35.835	31.643
			36.1433	37.947	33.682
SRResNet_bicx3_in3nf64nb16	3	RGB	31.019	32.442	28.499
			32.449	34.428	30.371
SRResNet_bicx4_in3nf64nb16	4	RGB	29.051	30.278	26.853
			30.486	32.180	28.645
SRResNet_bicx8_in3nf64nb16	8	RGB	25.429	25.357	23.348
			26.885	27.070	24.996
SRResNet_bicx2_in1nf64nb16	2	Y	35.870	37.864	33.581
SRResNet_bicx3_in1nf64nb16	3	Y	32.182	34.263	30.186
SRResNet_bicx4_in1nf64nb16	4	Y	30.224	32.0384	28.494
SRResNet_bicx8_in1nf64nb16	8	Y	26.660	26.621	24.804

¹ bic: MATLAB bicubic downsampling; in3: input has 3 channels; nf64: 64 feature maps; nb16: 16 residual blocks.

² DIV2K 0801 ~ 0900 validation images.

³ The first row is evaluated on RGB channels, while the secone row is evaluated on Y channel (of YCbCr).

⁴ (31.901, 29.711)