AI generated fetish images

Hmmm... How?

If you run SD at home as per the tutorial in the OP, there does not seem to be a restriction in place.
I think it only applies to those that can be used online (like in their Discord or on huggingface).

While you won't get highly detailed and accurate closeups of genitalia, the attempts of the AI to create them are not blocked or censored internally.

Yes, I tried (for science!), and no, I won't post the results (most of them are still garbage)

And as you can see in one of my previous posts, SD had no qualms with putting Emma Watson in a skintight catsuit.

It appears bondage related prompts work better with chains than with rope.
At least it sometimes results in proper posture.

[Like Ra]

Handcuffs work perfectly too.

If you want to dig a little deeper, there is now a free standalone version of another AI generator, called "Stable diffusion" you can run independently on your own PC.

It is VERY impressive, or at least it can be, if you experimented with it a bit and know what you're doing, see my first tries below. Unfortunately, it does not seem to be trained to be very ... kinky, yet.
  
Here's the blog post announcing the release with some examples.

Here's an aggregator for work created with the online version and the associated text prompts.
It's insane what this AI is capable of if fed the correct prompts and given potent hardware.

Here is a step-by-step tutorial on how to set it up in complete layman's terms.

Source: https://www.youtube.com/watch?v=ZCJpIDADzTI


Everything is explained start to finish, I just completed the process myself and I have no real programming experience beyond html/css.

I didn't see any requirements mentioned, but you will probably need an nVidia 20x0 series or higher with CUDA cores and at least 8 GB of VRAM (or you could get out of memory errors). My 2070S was working up quite a sweat and with 8 GB I could not do 1024 x 768 in size, so I stuck to 512 x 512.

A 24 GB 4090 looks quite tempting now.

There are some bumps in the tutorial video, I am outlining them here so you don't run into them as well.

You will run into an error. He corrects his file structure at around 9:30 after Anaconda throws the same error for him, just follow along and it will work - until you get another error because the video is also missing the crucial step to install Git (he forgot to uninstall it, so he will not get an error, but you will if you follow along don't have Git set up), a follow-up video describes how to:

Source: https://www.youtube.com/watch?v=lt9oDAvpG4I


Do this before you build the environment in Anaconda with the command
conda env create -f environment.yaml
or you will get another error and will have to fix it by following the second video completely (you will have to remove the environment).

I'll give it a go - I'll try it on ubuntu instead though... might be easier to run in the cloud also this way...   has anyone else managed to achieve this?

Some interesting... designs.

[Like Ra]

Once you find a base image you like, you can use that seed with some miniscule variation and just switch the outfit.

So you can redress the celebrities and anime girls 😋

Short trip report.

I futzed around with the inpainting feature most of the AIs offer.

Basically, you take a source image and mimic what you want the AI to do and describe it to the AI as best as possible.

Since the AIs currently are shamefully vanilla when it comes to kink (but you can bet your ass there are plenty of people working on changing that by training their own text inversion models - in theory I could do it too, if I had a GPU with more VRAM)

So I took one of my images, took a stock photo of rope and roughly laid it over the image, resembling a bondage scenario in the widest sense. Added a sketched head harness for shits and giggles. Unfortunately the AI has NO idea what to do with it and ignores it completely most of the time, and even if tries to do something with it, it's laughable. Eh, it was worth a shot.

   

Then I used this image and fed it into the AI as a source image, entered a verbose description and had it go for it- 32 times with a little variance parameter for some creative leeway.

Some of the better results below.

I imagine by preparing a source model with a better fitting pose and some tweaking of the rough "before" sketch this could produce interesting results. You'll note the "rope physics" of wrapping rather snugly around the body is fine, the rigging just needs some more work.

[Like Ra]

This is truly the stuff of nightmares!