AI Scaling Makes Classic Games More Beautiful

Pixel learning becomes HD graphics with comparatively little effort thanks to deep learning.

The modification “Enhanced Textures” for the role-playing classic Morrowind extrapolates the original textures of the game to four times the resolution with a neural network.

At its core, the AI ​​process is the same as that of the video counterfeiting AI deepfakes and other image manipulations : A scaled- optimized GAN network (ESRGAN) hallucinates details that match the texture and thus increases the resolution with additional information.

The low-resolution textures are enhanced by additional pre-trained neural networks.

A generator AI creates a texture proposal, a discriminator AI checks it for credibility. With each new attempt, the generator AI approaches the target of the discriminator AI. The ESRGAN network was pre-trained with high-resolution photos and their bicubically scaled down form.

The result is HD graphics that match the game world, which compared to manual post-processing of individual textures in Photoshop can be achieved without great effort and corresponds to the original style.

The code for the neural network is available free of charge . Instructions are here .