This artificial intelligence restores old photos with great quality

This artificial intelligence restores old photos with great quality

In most cases, old photos of our ancestors are usually in quite improvable condition. From scratches on the paper, through a high amount of noise, to quite poor colors and sharpness. But don’t worry, if you want to see their faces in high quality, this artificial intelligence helps you achieve it.

Using the power of the GFP-GAN model, the team behind Baseten has enabled a website where you can remove the passage of time in old photos. Thus, you will not only be able to digitize them and save them in a cloud storage service, but now you will also be able to give them some touch-ups that will increase their quality before uploading them.

This intelligence, in addition to improving sharpness and detail in the face, will also enhance the colors of the image. Likewise, removes paper crack lines and other blemishes common when a photo has passed through many hands, for a long time.

The results of this artificial intelligence are quite similar to another one that we already told you about at the time. However, that one made your old photos move, almost as if it were a video of your ancestors.

This is how the artificial intelligence of GFP-GAN works to restore old photos

GFP-GAN

GFP-GAN (Generative Facial Prior) was reported for the first time in a research carried out by Xintao Wang, Yu Li, Honglun Zhang and Ting Shan. In it they proposed GFP-GAN as a new architecture within the GAN model, which would allow rescaling the quality of human faces in photos that have been damaged by the passage of time. Together with the use of other models, it can offer surprising results.

This model is capable of repairing the flaws in an image in seconds. But, how exactly do you go about getting it?

First, a convolutional neural network (U-Net) uses the damaged photo to remove its degradations. This architecture segregates the image into two features that will allow you to work on it in a different way. To do this, they use StyleGAN, a network that allows generating faces with artificial intelligence and that has been associated with the creation of deepfakes.

Read:  This artificial intelligence can identify the sex and age of a baby

With StyleGAN, on the other hand, GFP-GAN can generate a new clean face from the data collected from the image. After going through various filters, layers and parameters that convert the values ​​to vectors, you can get a result almost identical to the original photo.

Yes indeed, this model still has some flaws. Since the artificial intelligence needs to be able to recognize the face in the image, it can generate a slightly different face if the image is in a very deteriorated state. The problem is in the inability of the model to recognize certain features, something quite acceptable if they have been compromised in the original photo.

How to restore your old photos easily

GFP-GAN artificial intelligence

Now that you know how it works, time to get down to business restoring those family photos. In case you don’t have any on hand, the website leaves you with some sample options. You can also turn to Google and search for old images, so you can test the capabilities of this artificial intelligence without much effort.

  1. Enter the website where the model is published and in operation.
  2. If you want to try some of the examples available on the page, just scroll down. Here you will see a collection of images. Tap on any and it will automatically be processed.
  3. If you want to upload your own photo, tap on the section Choose file…right next to the blue button Restore photo. Find the image in your storage, and then tap on the mentioned button to start the process.
  4. Now, you will see the original image on the left, and the enhanced version on the right after a few seconds of processing. To download it in full resolution, tap the download button. Download the restored imagejust below the modified version.
  5. Then right click on the open image and select Save Image As.