Battle of the AI apps: Photoshop vs GFP-GAN

A test comparing two photo restoration applications.

After reading a Medium post, “Repair Old Photographs with GFP-GAN Artificial Intelligence” by Ginger Liu, I decided to give the app a whirl. I wanted to compare the results between GFP-GAN and Photoshop's new Photo Restoration Neural Filter that shipped in Beta form with PS 2023 this October.

Here are the results of my test to restore a photograph from a Library of Congress public domain collection. Granted, the original image is not terribly damaged or fuzzy. I selected a photo with clear faces and lack of sharpness and contrast.

The original is a 4155 x 2337 pixel TIF file. I uploaded the file here and used the restoration app built with Baseten and powered by the GFP-GAN model. The process only took seconds and returned a PGN image that was resized to 2800 x 1574 pixels. Here is their restoration:

It looks sharper and smoother. Dust and scratches weren't fully removed. Next, I tried Photoshop 2023's (version 24.0.0 with Windows 10) new Photo Restoration Neural Filter. Here is a screenshot of the settings I used. I left the additional Adjustments at zero.

This is the result of the Neural Filter restoration, which I also saved and sized to match the GFP-GAN model:

Photoshop did a better job on reducing the scratches. The “Enhance face” dialed to 60 percent might have been too much for this particular image. Here are the two side-by-side, with the Neural Filter results on the left and GFP-GAN's results on the right:

What do you think? Which AI did it better?