This Version of GTA: San Andreas Was Supposed to Delight, and It Looks Like a Nightmare. Real AI Revolution Is Still a Long Way Off

Source: Niccyan

i

Generative AI is the hottest current technological topic. So it's no wonder that companies in the gaming industry are experimenting with it more and more often.

One of the ideas is to generate all graphics using AI models. It doesn't work out as hoped for, because they don't simulate a virtual world - they can't maintain a consistent image and everything falls apart too quickly. An alternative method, recently proposed by many enthusiasts of generative AI, is to run a game that would have very simple graphics, and then live enhance this image by AI to high detail.

Nightmare GTA: San Andreas

The latest example of what this could look like was posted on Niccyans YouTube channel and uses Grand Thef Auto: San Andreas.

If the author intended this material to increase player enthusiasm for generative AI, it probably had the opposite effect.

1. The Runway Gen-3 model used is nunable to maintain consistency even for a few seconds.
2. It constantly changes the details of the background and characters.
3. The protagonist of GTA:SA, CJ, has a different face in every shot, and in several shots his skin color even changes.

And it should be added that even these ugly images are not generated in real time.

Of course, generative AI is constantly being improved, but the fundamental aspects of this technology mean that it won't be able to replace traditional graphic engines anytime soon. Hallucinations are something that cannot be completely eliminated, because these models really do not understand what they generate, nor do they keep in memory models of the entire virtual world. In addition, even if we eliminate these defects, such technology would not allow for the creation of exactly the artistic style that the developers imagined.

Generative AI in games, but from a different perspective

This was seen in another recent example, in the form of Tencent's GameGen project, whose model was trained on footage from open-world games (without the consent of their creators, of course). Not only did it look hideous with numerous hallucinations, but its official page was quickly deleted afterwards. No wonder, as it was more of a show in the style of "how advanced we are and how we use AI", rather than something you could actually play. The generated images rather harmed the image of the Chinese giant than helped it.

A way to get around such problems is a recent experiment by members of the Google Research group, in which AI generated a playable version of Doom without any other engine. However, it was more of a cool trick than something that could be actually used in video games. The model was trained on the game itself.

Thus, even if we wanted to use such a method, we would first have to create the entire game, and then spend a lot of time and money training a model on it. In addition, even with such training material, this AI Doom was hallucinating quite often. As a result, developing such a game would be much more expensive, require more powerful equipment, and in the end, the quality would still be lower.

This does not mean that generative AI has no application in video games. Recently, Revolution Software boasted that it used such technology to speed up the creation of animations in Broken Sword: Shadow of the Templars - Reforged. However, the vision of all graphics being generated in real time by AI is not something that is likely to become a standard in the gaming industry any time soon.

Besides, even the idea of making the graphics more realistic can be implemented in a much better way. A few years ago, there was a lot of talk about a method of making GTA 5 photorealistic by live modifying the graphics using a rich database of city street images. However, this method required a lot of work, not just feeding the image into a ready-made AI model, so it's no wonder that it doesn't excite people who expect immediate results from artificial intelligence, without major costs.