Español 
English (UK) 
When I first started developing the idea for the game, I knew that the player should have a wide variety of different cars available and that they should be real. In fact, in the first videos I published, you can see a Koenigsegg and a Renault Clio.
I soon set a goal of having about 20 different cars, at a rate of 4 or 5 for each racing category. With that in mind, I started looking for fan-made models under CC-BY licenses on 3D modeling sites (which is where the two models mentioned above came from, and where I found a few more candidates for SoER)
However, as the basic engine evolved, I realized it was safer to avoid real-world cars to bypass potential licensing and trademark issues. This led me to the conclusion of using AI-generated models instead
The Process
I quickly started testing with prompts and different AI models. At that time, Hunyuan 3D 2 had just been released; it worked relatively well at generating 3D models if you provided it with several images (front, side, etc.). This also coincided with a new version of Gemini that produced good results generating images of fictional cars and, furthermore, was the only one I tested capable of generating three views of the same car in a single image (side, front, and rear).
From here, with the process devised, the Koenigsegg became the Laprini and the Clio was replaced by the Rapide R1. Both are fictional but have performance levels very similar to the previous ones.
Unfortunately, the quality of the created object (car) was a far cry from the previous ones, which meant the game's visual quality dropped significantly, given that the player's attention is primarily focused on the vehicles. I have to admit I felt a bit "down" when I integrated them into the game, but I immediately thought that new AI 3D object generation models would emerge and, indeed, that has happened—perhaps sooner and better than I expected.
Trellis 2
A few weeks ago, the Trellis 2 model was released, and I have to say it works wonderfully. The difference from Hunyuan is abysmal. The results at first glance are clearly superior, but it also provides transparency and even models the vehicle interiors. In other words, I can now generate better-looking cars with transparent windows that show their interior. You can compare for yourselves.
What hasn't been solved yet is creating a 3D model with separate parts (wheels, windows, etc.), as is common in fan-made models. It's true there are models that create objects taking their parts into account, but I haven't tried them and I don't think they generate textures. For now, what the technology behind Trellis offers suits me very well.
It's Not All About AI
I'm trying to be quite transparent about how I use AI in making the game. I'm aware that in some cases there can be some animosity towards the use of these kinds of technologies to create games. I simply believe it is a tool that is here to stay and, whether we like it or not, we have to learn to use it and live with it.
I'll say again what I mentioned some time ago: SoER is not (and will not be) one of those games created in a couple of hours by an AI. There is a lot of work behind all this, and the creation of the cars is a perfect example. Once the model is generated, I have to dive into Blender, set up an animation, and generate the images that will then be converted into sprites. In fact, I've had to learn how to use this 3D tool, at least the basics, to get all this done.
In the end, the process is complex and also involves researching models, prompt engineering, creating scripts in Python or Bash... It's quite a lot, but that's part of the fun. I think the result has improved significantly, and I'm very happy with how the new versions of the cars look.
By the way, I recently launched the game's page on Itch.io where you can find the playable demo with the previous models (Hunyuan 3D 2):
🎮 Download Demo on Itch.io