I’ve applied Neural Style Transfer to Return To Castle Wolfstein

TLDR: It looks great but not playable.

It’s aesthetically pleasing but screen is hard to read and you can miss some important details (e.g. bullet shoots or mission objectives) that hard to read due to the style is being applied. I’ve converted some torch models to the CoreML format and also trained mine own with Torch and ML Create. Model performance varies because each model’s configuration and architecture differ. Converted Torch models ~12 fps, my own trained torch models ~20 fps, Apple’s native CoreML models are fast ~60-70 fps.

Demo is linked down below 👇🏻