Grand Theft Auto V Looks Almost Photorealistic Thanks To Machine Learning

While we’ve seen loads of examples of AI making games look better over the last couple of years, I’ve never seen anything like this approach, that’s able to take Grand Theft Auto V and through the power of magic make it look so damn real.

Advertisement

This is a project by Stephan R. Richter, Hassan Abu AlHaija and Vladlen Koltun at Cornell University, culminating in a paper called Enhancing Photorealism Enhancement. Both the paper and the accompanying video get pretty heavy on technical details, so here’s the basic summary of what they’re doing:

We present an approach to enhancing the realism of synthetic images. The images are enhanced by a convolutional network that leverages intermediate representations produced by conventional rendering pipelines. The network is trained via a novel adversarial objective, which provides strong supervision at multiple perceptual levels. We analyze scene layout distributions in commonly used datasets and find that they differ in important ways. We hypothesize that this is one of the causes of strong artifacts that can be observed in the results of many prior methods. To address this we propose a new strategy for sampling image patches during training. We also introduce multiple architectural improvements in the deep network modules used for photorealism enhancement. We confirm the benefits of our contributions in controlled experiments and report substantial gains in stability and realism in comparison to recent image-to-image translation methods and a variety of other baselines.

In short: machines can only work with what they’re told, so by improving the quality of what’s being fed to them, and the way they’re being fed, the results can be drastically improved. Like this:

If you want to get really into it, even more so than in the video above, you can read the whole paper here.

.

Luke Plunkett is a Senior Editor based in Canberra, Australia. He has written a book on cosplay, designed a game about airplanes, and also runs cosplay.kotaku.com.

DISCUSSION

NoelVeiga

Hey.

Hey, everyone?

This isn’t about “making the game look better”.

This is a machine learning technique that looks at a library of pictures, then looks at a synthetic image and successfully makes one look like the other. It’s not about whether you like what comes out the other end, the insane part is that a computer looked at the game, understood what each object is *just by looking at the rendered frame* and then modified it to look like a reference video.

That’s goddamn insane.

And it goes by in a flash, so I’d like to see the render times on this, but they do say this is done at “interactive speeds”. If that’s true the options this unlocks are very interesting. The whole idea is not that much different from DLSS, where they are using the extra information a computer renders when they make your videogame pretty pictures to upscale them from a lower resolution. Now these guys are saying they may be able to do the same with other things besides resolution. Extrapolating the possibilities this could lead to a future where a developer just places a cylinder on a scene and tells an AI to draw a tree on top of it. Instead of having to calculate motion and alpha transparency on thousands of individual leaves you just let your GPU fundamentally dream up what a tree looks like and spit it out in the right place. You could make a ground truth 3D image with no shaders and instead just flag what type of material each of them is based on reference and have an AI “paint” the textures on top. You could feed it very low-poly models and let it add all the fine detail for big performance boosts. Doesn’t even have to be photorrealistic. Had they trained this with Renaissance paintings the output would look like one. Assuming they found enough XVI century paintings of station wagons, I suppose.

At the pace these techniques are moving we may be about to see the next big leap in real time graphics since we figured out how to hardware accelerate drawing polygons. This is not about whether you think the pictures in the tech demo look “too grey”.