Making a game using only AI

So I have set myself a challenge to make an entire game using AI - those that know my games will appreciate the gravity of the task.

Step one is to train AI on existing graphics to get a baseline 500 or so assets. Then to modify and colour them (as they seem to turn out rather plain). Then use these assets to generate the 1.5k assets I’ll need.

The setting is slightly futuristic. Step 1 is currently under way and here is the first pass results. So far so good as a base

Will update this thread as things progress if it proves successful or not.

3 Likes

That’s cool! Interested to follow the progress on this!

Wish you good luck with this challenge. Sounds interesting. Will be checking this thread for the updates. And maybe very soon we will try your new game - AI Designer City :wink:

Can you also share the tools you’ll be using in the process?

So a small update…

Image 1 is the first gen (made from existing assets).
Image 2 is a mix of a tree layer and the important bit of the first gen (this can be rough as we will regen it anyway).
Image 3 is a refinement gen (much lower variance) to get the final asset.

1,499 to go…

Could you share your prompts? It would be very interesting to see how you coax the AI from 1 to 2 to 3. (Also, which generator are you using? Midjourney, Dall*E, etc.)

As I said already, I am using my own assets as the source so there are no prompts. To clarify what I wrote (sorry if this wasn’t clear?) - 1 is the initial gen from my existing asset base, 2 is not a gen step so there is no coaxing required, 3 is the final gen.

The tricky part is making a tree layer that gens nicely to get the final output I want. This part has taken a couple of days!

Now the process is simple and repeatable and step 1 to step 3 takes less than 5 minutes.

I’m not using any generic AI as they are terrible for anything other then conceptual art (and they are completely hopeless at isometric with transparency).

I understood that you had trained the AI on your own assets. I realize now that because I haven’t worked much with image AI, only text-based AI that I was using the wrong nomenclature. What I was curious to know is how did you guide the AI towards what you wanted? I mean, the AI doesn’t know what trees are, so how did you get it to improve the tree layer? Or are you taking step 2 and then manually turning it into the final product?
And then, also, if you’re not using a generic AI, what AI are you using?

As previously stated, step 3 is pure AI (based on the image I made in step 2). I must admit I don’t understand how it works either… but it does.

I’m using an architectural AI so it understands what I want. Generic AI fails badly with this.