AI Tools and Other Helpful AI things to help Game Creation

[Momo1521]

I want to say that I may be behind the times in knowledge about using AI as I have not even played with ChatGPT but after I saw that Love Season: Farmer's Dreams used AI in such a way I thought to create a thread to gather those tools and other suggestions.

Now this is not a thread to hate on the games that suddenly started to use AI to generate the graphics (such as Love Season: Farmer's Dreams) and that's not what you pay him for as a patreon or whatever. I am sure AI will eventually catch up and it will be indistinguishable. This thread is help gather all the AI tools, tips, tricks, and tutorials so that someone who never created a game before might be about to or to even help some of these abandoned games get completed. For example I have always wanted to create a game and have a story in mind but am engineer and do not have the arts mindset/vision and patience to create the graphics, background, poses etc even though I tried and spent a few thousand on DAZ.

So if everyone could contribute to post AI tools, tips, tricks, and tutorials on everything need to complete a game that would be great. Off the top of my head I can think of the following items:
- AI Image Creation
--Which ones are good
--Which ones allow the type of 18+ image generation need for games like we need here
--How to use them
--How to get the AI to base the images on DAZ models or other created images like in the case of Love Season: Farmer's Dreams (game started non AI but latest update used AI)
--How to get AI to get the images to be based off celebrities or real people
--How to get the AI to get the real people based image to look DAZ like (or other took like unity) so that it can not look out of place against other models (maybe even the AI porn tools that allow you to create nude real looking images based on their images)
--How to get AI to make the background and location graphics
-AI that helps with script generation and how to use them/get them (probably ChatGPT but idk there could be others or variants that are specialized in doing that)
-AI that helps with code generation on the different engines ex. Unity, Renpy, RPGM
-AI that helps with map generation on RPGM

Any other suggestions.

 

[anne O'nymous]

-AI that helps with code generation on the different engines ex. Unity, Renpy, RPGM

They don't exist.

AI do not think, they guess. They do not understand, they assume. Therefore, when it come to code, they will just take what they think correspond to your needs, and for this they'll follow a process that, in the end, is as accurate than you asking the same question to a search engine. They'll just dedicate a bit more time to this, trying to cross the references before it give you an answer.
If the question is about a really specific point that is hard to deal with, the answer have chance to be accurate and working. But if it's something basic, you'll get what "looks like" the most frequent and assumed useful answer. Except that in those case the said "most frequent and assumed useful answer" is rarely working.

By example, recently I wanted to draw lines in Ren'Py. I knew that to do, the Bresenham line algorithm is 60 years old, but why loose time porting it to Python myself when there's surely hundreds of people who published their own port online ? So I jumped on the occasion to test AIs and asked ChatGPT.
The code was working, but having used the algorithm in the past with other languages, I felt that something was odd. And I was right, the algorithm was only working for one quadrant (x < x1 and y < y1), the three other quadrants just give totally random result.
Searching deeper, it's only on the third page of Google result, that there's an effectively working port of the algorithm. All the other result, whatever how high ranked can be the sites hosting them, and whatever how high ranked can be the algorithm itself, give answers that are broken in a way or another...


It's not like writing, that you can correct by yourself if the AI give you inconsistent result. It's not like drawing, where you can see that the result is weird, and that you can either redo or edit. But code need that you understand what it's doing. Else you end with something that seem to works, will works in perfect condition, then will blow up in your face the day the condition will stop to be perfect.
Except that 99% of the people who will rely on AIs for their code will do it due to their lack of knowledge. So, they'll not see that the code is wrong, and the day it will explode they'll not know how to fix it. They'll have a months old game that they can't continue, because it's to late to remove that piece of code, and they can't fix it.

 

[Momo1521]

They don't exist.

AI do not think, they guess. They do not understand, they assume. Therefore, when it come to code, they will just take what they think correspond to your needs, and for this they'll follow a process that, in the end, is as accurate than you asking the same question to a search engine. They'll just dedicate a bit more time to this, trying to cross the references before it give you an answer.
If the question is about a really specific point that is hard to deal with, the answer have chance to be accurate and working. But if it's something basic, you'll get what "looks like" the most frequent and assumed useful answer. Except that in those case the said "most frequent and assumed useful answer" is rarely working.

By example, recently I wanted to draw lines in Ren'Py. I knew that to do, the Bresenham line algorithm is 60 years old, but why loose time porting it to Python myself when there's surely hundreds of people who published their own port online ? So I jumped on the occasion to test AIs and asked ChatGPT.
The code was working, but having used the algorithm in the past with other languages, I felt that something was odd. And I was right, the algorithm was only working for one quadrant (x < x1 and y < y1), the three other quadrants just give totally random result.
Searching deeper, it's only on the third page of Google result, that there's an effectively working port of the algorithm. All the other result, whatever how high ranked can be the sites hosting them, and whatever how high ranked can be the algorithm itself, give answers that are broken in a way or another...


It's not like writing, that you can correct by yourself if the AI give you inconsistent result. It's not like drawing, where you can see that the result is weird, and that you can either redo or edit. But code need that you understand what it's doing. Else you end with something that seem to works, will works in perfect condition, then will blow up in your face the day the condition will stop to be perfect.
Except that 99% of the people who will rely on AIs for their code will do it due to their lack of knowledge. So, they'll not see that the code is wrong, and the day it will explode they'll not know how to fix it. They'll have a months old game that they can't continue, because it's to late to remove that piece of code, and they can't fix it.

Thanks for the reply and taking the time to do a detailed response.

 

[belladore]

- AI Image Creation

I would recommend Stable Diffusion. Particularly derivatives of the 1.5 release, in combination with other resources, particularly LoRAs and ControlNet. For UI I would recommend either ComfyUI or a1111 webui. You can find the resources on Civitai and Huggingface. To get started, you should go through some tutorials, for example, Aitrepreneur's youtube videos.

 

[theporntux]

Longtime lurker, I actually study these sorts of models so I'll have a go at this. Would love to see more AI-driven development and art here at f95.

--Which ones are good

StableDiffusion 1.5 as the longest lived model is the most well supported (more inversions, LoRA, checkpoints).
StableDiffusionXL is better at generating up to 1024x1024 with good quality, and is generally easier to prompt.
StableDiffusionXL_Turbo is the fastest generator and has the same architecture as StableDiffusionXL, so it accepts LoRAs from StableDiffusion.

--Which ones allow the type of 18+ image generation need for games like we need here

The "official" models are SFW only, but merges and checkpoints trained on NSFW images are also available through communities like civit.ai

--How to use them

Get a beefy GPU (the more VRAM, the better. if you want to train your own LoRA, you'll benefit from a 3090 or 4090).
Install CUDA
Install StableDiffusion WebUI from Automatic1111
Download some pretrained models into the `models/stable diffusion` folder
Get good at prompting

--How to get the AI to base the images on DAZ models or other created images like in the case of Love Season: Farmer's Dreams (game started non AI but latest update used AI)
--How to get AI to get the images to be based off celebrities or real people
--How to get the AI to get the real people based image to look DAZ like (or other took like unity) so that it can not look out of place against other models (maybe even the AI porn tools that allow you to create nude real looking images based on their images)

I'm answering these in bulk as they all have the same answer: LoRA (Low Rank Adaptation). It's a method used in StableDiffusion (and other types of models) to adapt the outputs to be closer to a set of reference images. This requires a lot of GPU horsepower and memory.

--How to get AI to make the background and location graphics

Pretty easy, generate the background first (describe the background only in the prompt), then generate each character image with `(solid green screen background:1.5)` in the prompt, and use a color mask (green screen) in an image editor. This is one thing I'd really like to wrap into a workflow.

Alternatively, you can use img2img inpaint sketch mode, load the background image generated earlier, and draw a vague area where the character should appear. This has a tendency to make some of the background "warp" around the character in between generations, though, as it will slightly redo the background within your sketch.

-AI that helps with script generation and how to use them/get them (probably ChatGPT but idk there could be others or variants that are specialized in doing that)

ChatGPT is censored heavily. With a medium-sized GPU, you can run TheBloke/Wizard-Vicuna-13B-Uncensored-GGML (available at huggingface.co) which is particularly good at writing smut.

-AI that helps with code generation on the different engines ex. Unity, Renpy, RPGM

GPT-4 (and GitHub Copilot)/CodeLLama2

-AI that helps with map generation on RPGM

Don't know this one; what format are RPGM maps in?

 

[michel.gellar7]

Great subject!
Until know, it's really hard to generate an image based on a same model, same environnement. As if the prompt regenerate new parameters each time.
Would be great to define and block environnements and characters parameters, so we only change the settings.

This way, once we created the new characters and environnements, we just need to prompt the situations.
That would be awesome! Maybe aven a DAZ integration at some point for even better quality!

 

[belladore]

Yep, consistency is hard to achieve. There's ControlNet, LoRAs, IPAdapter, FaceId, and a bunch of other methods to try to achieve that consistency, but it's hard.