AI Face Restoration using GFPGAN on Vultr Cloud GPU

Updated on August 21, 2023


Generative Facial Prior-Generative Adversarial Network (GFPGAN) is a blind face restoration model used to restore and improve images. The model works with multiple image types such as old images, low-resolution, noisy, blurry, and AI-generated images. Generative Facial Prior (GFP) handles the face restoration process through novel channel-split spatial feature transform layers. This enables a balance between realness and fidelity on generated images.

Sample Image

This article explains how to perform AI Face Restoration using GFPGAN on a Vultr Cloud GPU server. You are to restore a single image from a base URL, generate, and export improved images to compare the AI face restoration result.


Before you begin:

Set Up the Server Environment

In this section, set up the server environment and install the necessary packages to run the GFPGAN model as described below.

  1. Create a new directory to store generated images

     $ mkdir -p /home/user/output_images
  2. Install PyTorch

     $ pip3 install torch torchvision --index-url

    The above command installs PyTorch with pre-built CUDA 11.8 libraries. torch and torchvision offer a platform to develop and train deep learning models, along with tools and pre-trained models for Computer Vision tasks to speed up the process of building vision-based applications. To download the latest version, visit the PyTorch Documentation.

  3. Install Jupyter Notebook

     $ pip3 install notebook
  4. Allow the Jupyter Notebook port 8888 through the firewall

     $ sudo ufw allow 8888

    By default, UFW is active on Vultr servers. The above command allows incoming connections to the Jupyter Notebook port 8888

  5. Launch Jupyter Notebook

     $ jupyter notebook --ip= 

    The above command starts Jupyter Notebook and listens for incoming connections on all server interfaces with the directive

    When successful, a random token displays in your command output as below:

     [I 2023-08-10 12:57:52.455 ServerApp] Jupyter Server 2.7.0 is running at:
     [I 2023-08-10 12:57:52.455 ServerApp] http://HOSTNAME:8888/tree?token=73631c92ba278d265aedeb3b199bd4d48e5ef5b2eed0ae06
     [I 2023-08-10 12:57:52.455 ServerApp]
     [I 2023-08-10 12:57:52.455 ServerApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation).
     [W 2023-08-10 12:57:52.458 ServerApp] No web browser found: Error('could not locate runnable browser').
     [C 2023-08-10 12:57:52.458 ServerApp] 
         To access the server, open this file in a browser:
         Or copy and paste one of these URLs:

    If you receive an error when starting Jupyter Notebook. End your SSH connection, and start a new session to activate Jupyter Notebook on your server.

  6. Using a web browser of your choice, append the generated token to your Server IP to access Jupyter Notebook

  7. Within the Jupyter interface, click New, select Notebook from the dropdown list and create a new Python3 Kernel notebook.

    Create a new Notebook

Install Model Dependencies

In this section, use Jupyter Notebook to install necessary GFPGAN model dependencies as described in the steps below.

  1. In a new notebook cell, install the required model libraries

     !pip install realesrgan gfpgan basicsr
  2. Click the run button on the main menu bar or press Ctrl + Enter to execute the above command

  3. Import the model libraries

     from gfpgan import GFPGANer
     from basicsr.archs.rrdbnet_arch import RRDBNet
     from realesrgan import RealESRGANer
     from IPython.display import display, Image
     import requests
     import numpy as np
     import cv2
     import os

    Below is what each imported library does:

    • GFPGANer: is a class within the gfpgan package that used to perform the actual GFPGAN-based image super-resolution
    • RRDBNet: (Residual-in-Residual Dense Block Network) is a deep neural network architecture that's used in super-resolution algorithms and it's utilized by the GFPGAN model
    • RealESRGANer: Handles the image background information
    • display: allows you to display objects in the Jupyter Notebook output cell instead of a separate file or link
    • Image: allows you to create an Image object from various sources, such as a path, a URL, or raw image data
    • requests: Handles HTTP requests when fetching an image
    • numpy: Handles numerical computations with support for arrays and matrices
    • cv2: Performs computer vision tasks, such as image and video processing, object detection, and more.
    • os: creates a link to the operating system and handles operations related to the file system, environment variables, and other OS-specific tasks.
  4. Run the command to import all required model libraries

Set Up the Model

Initialise and set up the GFPGAN model in your Jupyter Notebook file as described in the steps below.

  1. Define Checkpoints

     arch = 'clean'
     model_name = 'GFPGANv1.4'
     gfpgan_checkpoint = ''
     realersgan_checkpoint = ''

    The checkpoints defined in the above command are:

    • arch: As set to clean, it's the clean version without custom-compiled CUDA extensions used in StyleGAN2
    • model_name, gfpgan_checkpoint, realersgan_checkpoint: Host a URL pointing to a checkpoint file for the GFPGAN and Real-ESRGAN model. The URL consists of .pth files used to save and load the latest model weights

    To get the latest version and update the above checkpoints, visit the official GFPGAN releases page.

  2. Define RRDBNet for background enhancements

     rrdbnet = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2)

    The following are the defined parameters:

    • num_in_ch: Specifies the number of input channels for the neural network. num_in_ch=3 means that the network expects input images with three color channels, and represent the Red, Green, and Blue (RGB) channels
    • num_out_ch: Specifies the number of neural network output channels. Just like num_in_ch, the num_out_ch=3 declaration represents that the network is to produce output images using RGB
    • num_feat: Sets the number of feature channels or filters used in the model to determine the depth or complexity of the network
    • num_block: Sets the number of RRDBs in the network. Increasing this value can make the network deeper and more powerful
    • num_grow_ch: Affects how the network scales and transforms information
    • scale: Sets the upscaling factor that the network can handle. scale=2 represents that the network can upscale the input images by a factor of 2
  3. Define the RealESRGANer object for background enhancements

     bg_upsampler = RealESRGANer(

    The above command parameters handle the following functions:

    • model_path: Specifies the path to the checkpoint file .pth that contains the pre-trained weights and parameters of the Real-ESRGAN model you defined earlier
    • model: Sets an instance of the RRDBNet object that represents the architecture used by the Real-ESRGAN algorithm.
    • tile: Specifies the size of tiles the input image divides to during the upsampling process. Splitting the image into tiles helps in handling large images that might not fit into memory in a batch. tile=400 divides the image into tile sizes of 400x400 pixels
    • tile_pad: Determines the number of pixels added as padding around each tile before the upsampling process begins.
    • pre_pad: Specifies additional padding added to the input image before the upsampling process starts
    • half=True: Performs upsampling functions using the half-precision floating-point arithmetic. Using False performs the upsampling using full-precision floating-point arithmetic
  4. Define the GFPGANer object for face restoration

     restorer = GFPGANer(

    The above parameters function as below:

    • model_path: Takes in the latest gfpgan_checkpoint version you defined earlier
    • upscale: Specifies the upscaling factor for the super-resolution. It represents how much the resolution of the input image should increase
    • arch: Takes in the latest arch version which you defined earlier
    • channel_multiplier: Controls the number of filters used in the GFPGAN model. Increasing this value makes the model better but also increases the GPU memory usage
  5. Create a function to retrieve the image

     def download_image(url):
         response = requests.get(url)
         image_array = np.frombuffer(response.content, dtype=np.uint8)
         image = cv2.imdecode(image_array, cv2.IMREAD_COLOR)
         return image

    The above code declares a function named download_image() with one argument which hosts an image URL. The declarations inside the download_image() function include:

    • response: Uses the requests.get() function to send an HTTP GET request to the provided URL to download the image from the URL
    • image_array: Stores the binary image data in the response.content attribute. It then uses NumPy's np.frombuffer() function to convert the binary data into a NumPy array of unsigned 8-bit integers (dtype=np.uint8). The array represents the image data in raw pixel format
    • image: Stores the cv2.imdecode function to decode raw image data stored in image_array. cv2.IMREAD_COLOR, and specifies that the image should be in color mode, meaning it's a 3-channel image (RGB)
  6. Run the function to prepare the model to import an image URL as the input and perform face restoration tasks

Run the Model

To run the GFPGAN model, import an image from a direct URL with a readable file extension such as .jpg, or .png. When ready, load the image as described in the following steps

  1. Load the image using the download_image() function

     image = download_image('')

    For a quick addition to the model, upload your image to a public image hosting platform such as imgur and copy the direct image address to import into the function.

  2. Load the image in the restorer function you defined earlier

     cropped_faces, restored_faces, restored_img = restorer.enhance(image)

    The restorer function breaks the image and stores it in three data types as below:

    • cropped_faces: Consists of cropped faces extracted from the input image
    • restored_faces: Consists of cropped faces when the restoration is complete
    • restored_img: Consists of the fully restored image

View the Generated Output Image

  1. Run the following function to load the processed image. Replace .jpg with your actual source image extension

     def cv2_imshow(image):
         _, encoded_image = cv2.imencode('.jpg', image)

    The above code creates the function cv2_imshow() with a single argument that consists of the image URL. The remaining parameters are:

    • _, encoded_image = cv2.imencode('.jpg', image): Uses cv2.imencode from the OpenCV library to encode the input image in a format defined by the user. The function returns two values, a boolean value that shows whether the encoding is successful, and encoded image data in the form of a NumPy array
    • display(Image(data=encoded_image.tobytes(): Uses IPython.display.Image to create an image object from the encoded image data and displays it in the output cell of the Jupyter Notebook using display(). The tobytes() method converts the NumPy array containing the encoded image into a bytes object for the Image class.
  2. To view the final images, run the following commands. Replace i with your desired array index value such as 0,1, or 2 depending on the number of faces in your image

     #display original cropped faces
     #display restored cropped faces
     #display restored image

    As applied in the above command, faces in the image are available in form of an array. Therefore every face has an index value (0,1,2,3 and so on). Set the number i depending on the face you want to view and the total number of faces in the image.

  3. To view the final generated image, run the following command


    Verify that the generated image contains notable improvements as compared to your original image.

Save Generated Images

  1. To save the final generated image to the output directory you created earlier, define the directory and filename as below

     output_directory = "/home/user/output_images/"
     output_filename = "image.jpg"
     output_filepath = os.path.join(output_directory, output_filename)
     if not os.path.exists(output_directory):

    The above code block consists of:

    • output_directory: The location to store your final image
    • output_filename: The name to save your exported image as
    • output_filepath: Consists of The os.path.join() function from the os module used to combine the output_directory and output_filename into a single string representing the complete path of the file to store the restored image
    • os.path.exists: Checks if the output path exists, else, it creates the path.
  2. Save the image

     cv2.imwrite(output_filepath, restored_img)



    The above command returns True when the image saves successfully to the directory, and below is what the function does:

    • cv2.imwrite(): Consists of a function from the cv2 library used to save the restored image to the specified output_filepath. You can replace restored_img with either cropped_faces[i], restored_faces[i], or keep restored_img depending on the target image you intend to save.
  3. To download a copy of the generated images, launch a new terminal session and establish an SFTP connection to the server

     $ sftp user@SERVER-IP
  4. When logged in, Switch to the output_images directory in your user home

     sftp> cd output_images
  5. View available files in the directory

     sftp> ls
  6. Download an image file of your choice. For example, to download the image.jpg file, run the following command

     sftp> get image.jpeg
  7. When the download is complete, close the SFTP connection

     sftp> exit
  8. Using your local computer's file explorer, find the image in your terminal's working directory

Memory Usage

In your Jupyter Notebook session, run the following command to view the GPU memory usage statistics


In Half-bit, your output should look like the one below:

| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|    0    0    0      50737      C   /usr/bin/python3                 3252MiB |

As displayed in the output, with half-bit precision, the model uses about 3 GB of VRAM.

With full-bit precision:

| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|    0    0    0      50855      C   /usr/bin/python3                 4492MiB |

In full-bit precision, the model uses 4.5 GB of VRAM.


In this article, you implemented the GFPGAN model on a Vultr Cloud GPU server. You prepared the server, installed dependency libraries, and used the model through Jupyter Notebook. For more information on how to use GFPGAN, visit the official model repository.

Next Steps

To implement more solutions on your Vultr Cloud GPU server, visit the following resources: