If you have already generated enough art with artificial intelligence, why not use the same technology to restore it? Think of all those personal photos that were damaged by the passage of time, or that suffered other forms of deterioration (heat, humidity, fungi, and more). The Baseten GFP-GAN system aims to offer a solution to these problems by restoring old photos in a single click, without additional user intervention.
We all know that Adobe Photoshop can do incredible things in the hands of a professional. Time seems to go backwards on some images, removing errors, stains, tears and folds. However, the general public wants access to that ‘magic’ without relying on expensive tools, and that’s where artificial intelligence comes in. One of the most interesting projects is Baseten GFP-GAN, so we decided to try it.
How to restore old photos with Baseten GFP-GAN
There are no advanced settings in its interface: Load the photo you want to restore, and click Restore Photo. If you don’t want to upload your own material, you can use one of the sample images as a ‘demo’. In most cases, what Baseten GFP-GAN does is remove the superficial deterioration of the photo, and optimize its definition with what appears to be a sharpening phase, especially on human faces. At the same time, the algorithm presents ‘its’ interpretation of the image, and there is a possibility that some people may fall a little into the ‘Uncanny Valley’.
Unfortunately, we have not obtained good results with photographs that show extreme damage. The algorithm tries to correct and optimize other aspects, but ends up ignoring the most important ‘wounds’. Needless to say, the model will improve with additional training, and exposure to a much larger number of images. The current version of GFP-GAN is at 1.3.4, and being an open source project, I imagine it will receive several updates in the coming months.
Beyond its errors and successes, I recommend that you do several tests of your own with Baseten GFP-GAN. Maybe the algorithm will get along much better with your own photos, presenting superior and cleaner results. Link below!
Official site: Click here
