If you figure out the right prompts and the right model it feels like you could get a pipeline going with SD. You still might want to hire someone to turn those into actual assets, but now someone can create maybe 5 assets in an hour instead of one in two hours. If you get fancy you might be able to use things like inpainting to create tiles that blend nicely.
If you're going to care about metrics like "asset creation per hour," there already exist countless ready-made assets online, many even free to use. If you care about quality over quantity, then the value of an artist spending two hours on a single asset is obvious.
There is no aspect of asset creation where using AI to generate assets results in a better quality game versus human artists. Using artists to "improve" AI generated content is counteproductive because of the lack of direct control over what the model generates, something you don't need to deal with if a human is creating the art to begin with.
There are already tools available that let you create tiles that blend nicely. You don't need AI for any of this.
AI would be useful automating tedious and time consuming tasks. I've seen many artists say they could use an AI tool for automatically unwrapping UVs, generating spritesheets, JSON files, etc.. but the only thing AI ever seems to be used for is automating the creative process itself, and the results are at best technically competent (which is a result of simulating the work of human artists in the dataset) but never innovative or creatively compelling.
This is what irritates me about the current generation of AI. It is designed to give the appearance of "making humans unnecessary", when people mostly need help with drudgery and busywork.
