Economics of Civitai

This is one of four images I created on Civitai for just 33 yellow buzz, which means it cost me just 8 buzz to create this image, which is unfortunately useless, as were the other three images. See, I only wanted one character in the image, but I made the mistake of describing her as "a naked harem slave" and the model focussed on that word "harem" which generally means several naked women. It gave me four of them. And since I had described the slave's posture carefully and thoroughly, the model thoughtfully distributed the arms and legs among the four harem slaves. Plus whatever fraction of a slave that the indescribable lumps of flesh scattered here and there with what I think we can all agree is marvelous creativity displayed by an LLM that has NO idea how human anatomy works, was. Maybe two thirds?

 Let's talk economics. To make NSFW images on Civitai, you need yellow buzz. You can buy yellow buzz at the rate of 1,000 yellow buzz for $1.00 USD. I buy yellow buzz in ten dollar increments, and 10,000 yellow buzz lasts me well over a month ordinarily.  It's not expensive for doing images (videos are pricier). 

Typically the way it works with the inexpensive image generators is, you create your prompt (or download your starting image if you are doing image to image generation like I did in my previous post with the Second Life avatars) and you ask for four iterations, generally at a cost of between 16 and 40 yellow buzz, mostly around 30 buzz. This gives you four varying images, some of which may work well, some of which may not. Most of the time the first iteration is not particularly successful, and another iteration may be called for. Even when you get a usable image, that doesn't mean you're finished most of the time.

Very frequently you get malformed hands/feet/faces or you get weird objects where they don't belong, or you get deformed genitals or ropes and chains going where they should not go, bonds not working, and ball gags... don't get me started on ball gags! NONE of the cheap models get ball gags right, most of the time.

Early iterations can be problematical.

There are fixes for all of these issues. There are programs called loras that work with the models (the programs that create the entire image) to solve specific problems. For examples, there are loras designed to create realistic ballgags, and loras that allow you to pose the images in various kinds of bondage and loras that create characters with certain traits, etc. The loras and models are free to use, but to see if/how they'll work means you have to create the images on Civitai and that takes yellow buzz ka-ching!

And not only are there loras, there are also programs that can improve images. The costs can vary on these: there is "hires fix" one of my favorites that creates the image at a higher resolution and generally spruces it up and fixes minor details (but not deformed hands or anything like that, though it often helps with wonky eyes) but it costs about 16 buzz for a single iteration. (Fortunately, if you do four iterations, all of them look pretty much the same, so there's no point in doing more than one hires fix iteration.) There's also facefix, which replaces the face of your subject if it comes out wonky as they often do. Facefix is relatively cheap, it costs the same as a simple prompt iteration so it's not expensive, but it's very hit or miss so you tend to do four iterations when you do facefix, which works wonders sometimes and does nothing useful other times.

What's more, the cheap Civitai models (and just about every AI art program on the Web) are TERRIBLE at following prompts. Part of it is flaws in all LLM art programs. For example:

I described the male character as standing, clothed, dominant. The program  completely ignored the prompt and showed him naked, kneeling, bound, collared and chained, just like the slavegirl. LLMs like to use the same descriptors for all characters. Supposedly using the word BREAK and keeping your character descriptions entirely separate works, but lemme tell ya, it doesn't. Also, most of the time, Civitai's less expensive models get cocks wrong. They can't draw them at all, they make them look deformed, or they put weird plastic extensions on cocks that you didn't ask for (not kidding!).  

In addition, Civitai's cheaper models will fight like hell to keep from putting female models in some submissive poses. They'll make female characters kneel on their knees as shown above, no problem. But they will be DAMNED if they will show a female character with her hands tied behind their backs. Or kneeling with her head pressed against the ground and her ass hiked up in the air. (Do NOT use the words "doggie style" in your prompt because you'll probably get a furry image.)

Beyond the helper programs (I have no idea what they're called) there are sliders you can push: Denoise, which controls how closely your image follows the prompt (more about the prompt later), CFG which controls how closely your image follows the model/loras, and each lora has its own slider too.

And every time you run an iteration with the models and the loras and the sliders and the helper programs it costs more buzz. And you tend to do that a lot, because the dopamine response when you get a usable image out of all that mess is pretty good. And even if you don't get exactly what you want from an image, you sometimes get images that you can save and use for future projects.

So you can see how it's VERY easy to spend more money on the cheaper models, and have a great time doing it. I know I have.  But there's another expense as well: time. It takes a lot of time to generate those iterations, to write and rewrite prompts, to figure out which loras/sliders/helper programs to use.

The near one-and-done nature of Flux and the other programs are matched or even outmatched in price by the cheaper models, to get an often inferior image, and they take a LOT more time.

Mind you, you can still play with the sliders and the prompts with the more expensive models to get just the result you want, or some approximation of it, but you don't HAVE to in order to get a usable image. But it will cost more money. And you can save all that money and more importantly, time and use it for stuff like... writing fiction.

But there's one reason I stick to Civitai. And that's the lack of censorship. Every other AI art service I have tried, it feels like wrestling an octopus to create NSFW or even NSFW-adjacent artwork. On Civitai if I want a ballgag or a blowjob or even a woman dressed in a skimpy thong, I can ask for it, and Civitai will attempt to deliver it. When I wanted to get a ballgagged woman on Leonardo or Tensor Art (don't even talk to me about MageSpace!) I'd have to describe the woman as having a ball in her mouth, not a ball gag. Then I'd have to draw in the straps, if I got anything close to looking right. Sigh.

Civitai is the place to go if you want to do NSFW or even NSFW-adjacent stuff.

In my next post I'll talk more about the censorship that threatens to engulf the AI art world.

Civitai, Tensor AI, MageSpace, Leonardo AI: Blunderers All

WARNING: This image is safe for work in every respect yet still is an AI blunder. All the images below it are Not Safe For Work and also Not Safe For Life in some cases. You have been warned.

MageSpace: This is Schrodinger's blunder: it is and is not a blunder, both at the same time. In the sense that the artwork on this image is well executed with no glaring issues, it's not a blunder. The problem is that this is an image to image creation, and the image it was supposed to be based on was a lot like this image in every respect, except the woman in the foreground is naked and has her back turned to the viewer, mostly, as she watches the dirigible float past. So it is a blunder in that sense. This is very symptomatic of MageSpace, which HATES nudity. There are ways to get around it (image to image) but it's very difficult at best. Don't even think about generating sexy art with prompts on MageSpace.

Civitai: When Civitai goes bad, it goes BAAAAD. It gets naked women in bondage all wrong in so many ways. I forget what image this is based on, but whatever it was, it didn't look a THING like this THING. What I find particularly ironic here is that Civitai is particularly bad with portraying gags (most AI art programs are) but in this case it did a pretty good job of rendering the ball part of the ballgag but then rendered the face around it in the most horrifying manner possible.

Tensor AI: Tensor is censored like all the AI programs right now. It's very hard to get good, clean, naked image out of it. Monstrosities like Spiderwoman here are fairly common. One hopes things will get better soon, but who knows with the right wing in power and censoring every fucking thing they can?

Leonardo AI: Leonardo AI is just plain cruel. It'll get your hopes up by producing an image of a bonnie lass with a smiling face on a sofa by a river and then... and then... you look at her breasts and notice that she has no nipples. And you look at her belly and notice that she has two bellybuttons, neither of them anywhere close to where they should be, which is hidden behind her forearm. Speaking of hidden things, her left hand has COMPLETELY disappeared into the cushion, which doesn't look all THAT soft. But perhaps that's a blessing in disguise, because her right hand which IS visible has some very wonky fingers. But then you notice the absolute worst, which is that the shin and foot of her left leg have somehow merged with the sofa's curlicue gilt backing. It is amazing how Leonardo AI can take such a beautiful image and utterly ruin it with one awful detail after another.

All AI art programs appear to have their weaknesses. I'm sure the art creators who present flawless image after flawless image are chest deep in culled images of the sort I'm showing you, which they are not showing you because they want you to believe they are AI art whizbangs. That's the nature of AI art, you have to cull a lot of dross to get the good stuff, and if AI art is an indicator for AI in general, we don't have to worry about AI matching human intelligence anytime soon. Much less superintelligence.