The neural network has learned to turn a pixel image into a photorealistic

Recently showed us a remake of the classic shooter 1997 Quake II, where the main work of improvement was reduced to the integration of technology ray tracing, but this time in a Network there was something more interesting. Finally, the HD remake that we’ve all been waiting for! A Reddit user using a set of several neural networks was able to turn pixeltown face image of a space marine from the classic 1993 shooter Doom in an almost photorealistic image.

As the portal Futurism, as a result, the face of the famous “man of Doom” after all the manipulations looked rather like a muscular version of the actor Nathan Fillion.

The end result perfectly demonstrates the abilities of modern artificial intelligence algorithms adapted to recreate the images based on very low quality of the source material.

Create photorealistic images

As explained by the source, the final result is the product of several generative-adversarial networks (Generative Adversarial Networks, GAN). This machine learning algorithms, based on a combination of two neural networks, one of which generates samples and the other tries to distinguish correct (“true”) samples from wrong. Prospects for this technology a great multitude, as we wrote in one of our previous articles.

Enthusiast who created these images, missed the first sprite person “Doom guy” through several editing programs photo images (FaceApp, Waifu2x and GIMP). The result, though was significantly better than the original image, but still was too pixeltown.

Transition result. The face is visibly transformed, but still a fake

Further work was carried out using the developed by NVIDIA generative-adversarial network StyleGAN. It is able to generate individual (not only persons but also inanimate objects) that never existed and the moment it is one of the most powerful generative models-adversarial neural networks, which reveals impressive visible results. To consolidate the results he missed the face through StyleGAN again. Since the original face image had unrealistic proportions, final touches and “smoothing the corners” had to be manually.

After treatment with GAN technology, but to edit manually

The final result

Compare the original sprite of images and processed

To discuss the news in our Telegram chat.

Leave a Reply

Your email address will not be published. Required fields are marked *