In a demonstration this week, Intel showed off a new machine-learning technique that uses neural networks to transform games like Grand Theft Auto into something almost completely — almost eerily — photorealistic.
Just check out this side-by-side. Compared to the original, Intel Labs’ version provides some photorealistic flourishes like smoothing the road, adding green to the mountains, and simulating reflections off of passing cars.
To train the neural network, Intel Labs familiarized its AI with the Cityscapes Dataset which contains footage of real-life streets in urban Germany.
As the researchers note in their demonstration, when graphics are altered with machine-learning the product is usually “temporally unstable,” meaning the resulting in-game environments are peppered with artifacts and hallucinations that make gameplay more abstract, Dalí-esque compositions than photorealistic facsimiles.