GFPGAN on Ubuntu: setup and face restoration tests

Learn how the speaker runs GFPGAN 1.3 on Ubuntu with a CUDA GPU, then compares restored faces from sample images and heavily blurred photos.

Player not loading? Watch on YouTube

This Ubuntu walkthrough tests GFPGAN for face restoration using repository samples and the speaker's own blurred photos. Uploaded on October 9, 2023, it uses version 1.3 according to the speaker. His comments about the latest version refer to that demonstration, not the current release.

The setup overview covers cloning the Python repository, installing dependencies and downloading pretrained weights. The speaker has already completed those steps before running inference, so viewers get an explanation rather than a full installation demonstration. He uses a Conda environment to keep the work separate from his normal Ubuntu installation.

The inference example takes a folder of images and writes results into folders for comparisons, restored whole images, cropped restored faces and input face crops. The speaker runs it on a CUDA GPU. He suggests CPU execution may also work, but does not test it. He also describes Real-ESRGAN as the background enhancement component, while GFPGAN restores the faces.

The local AI trial then moves to more severely blurred inputs, including a photo with multiple faces. The speaker finds some outputs convincing and points out failures at higher blur levels. These examples show the limits of the demonstrated restoration results; they do not establish that generated facial details accurately recover the original face.