They have! There's a Stable-Diffusion-enhanced version of Zork[1] out there, and there are also extensions[2] for Oobabooga which facilitate dynamic adventures. Also, this point-and-click adventure game[3] is pretty fun if you can get it to run.
Doesn't work, SD's comprehension is too weak to make complex scenes.
Now DALLE-3 is good enough, problem is if it'll be cheap enough. 100% there are revolutionary new forms of adventure games being developed based on it.
Tangentially related: a friend made a Discord bot that automatically generates D&D items, NPCs, monsters, etc. with images and full descriptions. It uses neural.love and Poe for images and text generation, respectively.
>Finally, online text-based role-playing games brought to life!
Not yet, OpenAI is too preachy , if your adventure would bring you into a bar and try to drink something you will get a few paragraphs about "alcohol is bad", and is also very "child limited/targeted" so if a monster spawn , the hero will most likely befriend the monster will love and they will live happily ever after.
The open uncensored ones are still WIP last tiem I checked, they either forget the conversation from a few moments ago, or the training made them dumb.
If I am wrong then someone tell me where I can try a demo for such a text adventure that has memory and is not crippled like it targets "small american children".
The intent is more about avoiding legal and political risk. Even a grocery chain's LLM recipe generator gets bad press. I guess they need to warn you that drinking bleach is bad.
I normally test LLMs by providing a list of bizarre "weapons" and asking how I could use them to defeat an improbable beast.
It turns out that an enormous dust mite the size of a car needs a disclaimer when your weapons are a comb, a plastic horseshoe, an etch-a-sketch, a baseball glove, and a worn copy of the farmer's almanac. That being said, ChatGPT still tends to wins the creativity test.
You could use relatively low res images to generate "accurate" ascii for a place
It would be relatively lightweight but still "realistic"
I'm not well versed on Ascii to image conversion, but I guess you could use AI to segment and recognize objects too complex to recognize easily and adjust "image" generation accordingly
“You are standing in an open field west of a white house, with a boarded front door. There is a small mailbox here.”