Using AI generated art for placeholders or incidentals?

[Melkor99]

So, I've worked with some people who know Ren'py, and we've ever so often thought about making a game. Just one problem. Art costs. Good art... costs a lot. And with all the games out there, people generally want to see something before they pay you so you can pay for art.
Well, Novel AI and other AI art generators out there can produce a lot of art. Not... great art. Mind you, sometimes you have things like the woman's head turns into a squid or one hand has 27 fingers. But that's fixible, and a lot cheaper than good art.

So, would people consider it bad form to use AI generated art as a "placeholder" so you could play the game, and see the action, with the intention of replacing it, as time and money allowed, with better art?

 

[Friedrich Nietzsche]

AI art has one major problem, harmony.
And when you want to create some kind of game, you really need it.

 

[IndolentCat]

This is old but in case someone else has this question. This game uses AI on 95% of the VN backgrounds https://f95zone.to/threads/love-n-war-hero-by-chance-ii-v1-0-1-reborn-entertainment.13527. It has exactly that problem, sometimes you can clearly see differences on style/palette, among others. It looks like there was no direction, but I'm sure the game is doing good anyways as it is the third installment. As long as it is not for character art, a VN should be fine, especially if you blur it a little.

 

[lancelotdulak]

This is old but in case someone else has this question. This game uses AI on 95% of the VN backgrounds https://f95zone.to/threads/love-n-war-hero-by-chance-ii-v1-0-1-reborn-entertainment.13527. It has exactly that problem, sometimes you can clearly see differences on style/palette, among others. It looks like there was no direction, but I'm sure the game is doing good anyways as it is the third installment. As long as it is not for character art, a VN should be fine, especially if you blur it a little.

Dont blame the competance of the mechanic on his tools..

 

[lancelotdulak]

AI art has one major problem, harmony.
And when you want to create some kind of game, you really need it.

this is literally designed into the core of the tools.. you just have to actually know how to use the tools. You can literally require lense , color palettes, art styles..

 

[SenMizeri]

this is literally designed into the core of the tools.. you just have to actually know how to use the tools. You can literally require lense , color palettes, art styles..

Hiya, I'm not contesting, I hope you're right, but how do you effectively synchronise decent avatars, and expressions, as well as sex scenes and more?

Seems like it'll run into some issues with 'harmony'/'consistency' in places like that.

 

[Julie L Griffin]

Three are several answers. The first for a lot of VN's is come up with some room you like for a background and just stick with it. A lot of traditional VN's after all, only have four or five backgrounds. That's wasy.
For avatars, expressions, sex scenes? That's a bit more difficult. The first method is the simplest: set your prompt, set for batch count: 100, batch size 8, go for lunch, come back and go through your 800 images for which one comes closest.
Presuming your prompt is good and exact, you'll find at least a few that are 'close enough for government work'.

The most complex method is to create a LORA, TI, or embedding that will create your character. I tend to go with LORAs, as they are the best and TBH, not that much harder to make. You create a bunch of characters the old fashioned way by prompting, caption the best files and generate a LORA. To use one example, I was playing around, generated about four hundred images with a fairly vague prompt, captaoined them lazily and got this: (note the top grid I added "large breasts" which is why she's bigger there).


Now that was lazy, I didn't do a lot with the captions or really curated my data set, but that's still a LORA which is giving decentish looks. Note that the clothing and background is prompts--when I made the LORA the character was naked, against a gray background, so tghe model isn't predijuced to any given look. For a VN, you'd probably want to make several loras based on this, each on with a different clothing type, because otherwise you get variation on the clothes.

For expression? In painting, a lot of inpaintaing. Go in, mask the face or lips, and say: smile. Be prepared to do some work in photoshop/whatever you have to touch up. It'll never be as good as a skilled artist, but it probably is good enough that people aren't wanting to hunt you down in the street.

 

[SenMizeri]

Yeah, I am finding 'AI' art is fantastic for backgrounds because I am not an artist, but now I can create any background for the scenes I need. That is a given, and I'm not bothered what others think about AI, there simply isn't enough BG art around for new devs who can't draw.

What is a LORA or TI?

 

[Julie L Griffin]

okay, one of the issues with AI art is that models tend to vary. So LORAs and textual inversions are ways of convincing a model to do a more accurate job. A TI is essentially a fancy way of creating a "prompt" that is more accurate, without taking up a lot of tokens. It was one of hte first things to come out to try to make models more accurate, say, you want a particular look of a red head.

The problem is that it doesn't actually change the model itself, so if the model doesn't have enough reference, it wont' work. Good news, TI's range from 6KB to 50KB so they're small.

A LORA is sort of like a mini model that works along side your base model. They're bigger, up to 200MBs in size, but they work a lot better, and often can be used on a number of different models. You train them on images, and well, the above images are an example of what you can get. And you can modify them.
That charactger above was trained nude, so the model just uses her base body, if I wanted say to make a lora for a nurse uniform, I'd generate her in "a nurse uniform" (probably a lot more detail) and then train another LORA on that model, so I'd be able to say, oh, nursegirlxyz*, and it'd cough up a consistantish look.

*if you try to train LORAs its very important to do something like add XYZ to your trigger term, becuase you don't want the model to get confused, say your character is named sheila, but sheila already exists as a name in the model, so you use XYZsheila to avoid confusion.

 

[Gabaw]

Excellent information, Julie L Griffin , I learned a lot. The images are great and it's frightening how consistent they are. When you help the robots put us into the Matrix pods, please remember to put me in with Lois and not Meg, thank you in advance.

 

[DuniX]

It depends on the game.
For VNs you are probably better off using the tried and true Daz or Illusion renders as you would have control to set up your scenes and do things like sex animations.

Where AI CG works best is in games where the scenes are more one off and can be more random and replaceable.
So Sandbox/Management/Simulation style games that usually use random images could well use AI CG instead.
That and I am looking to see how well AI CG could be used for other elements like sprites for combat and enemies like in this game:
https://f95zone.to/threads/rogue-femme-v0-0-1-demo-banana-stroke.158174/

I am not sure how well AI CG can do Furries and Transformations but if they could it could be interesting to see what it can do in a game like Corruption of Champions or games with a lot of fantasy races and possibly breeding.

 

[Julie L Griffin]

The biggest failings, even with everything added, right now is interactivity. You may ask for: guy and gir fucking, and then literlaly have to generate hundreds of images until you come up with something usable. Good news, most AI rigs can do just that. Bad news, even with that, I expect some inpainting and tweking so that people don't wonder why7 the guys dong is emerging from his naval.