Jump to content
Search In
  • More options...
Find results that contain...
Find results in...
Existing user? Sign In

Sign In



Sign Up

Playing with AI generated fetish art


Recommended Posts

You have have heard of recent advances in AI image generation... it's kind of fun to play around with it to generate some fun images in our niche.

 

Some examples from just a little time using https://www.craiyon.com/

ChPHIM.jpg

GltPuf.jpg

YU6eKv.jpg

(the above was probably the best prompt I ever tried, often gives great results)

 

gNOPZS.jpg

 

cp7eds.jpg

 

 

T76G8Y.jpg

6HjKW9.jpg

 

EwcQiZ.jpg

 

(yeah the faces are still scary on these models, but they're getting better at it slowly)

Obviously it has limitations and it's very hit or miss, having to play around with different words to see what you get, but that's the fun of it. Try it out and post your best results. Creative prompts sometimes give really great (or hilarious) results.

 

Link to comment

QqcFgz.jpg

The fun of this is realizing that one day—probably not too far away—we’ll be able to ask for any image and have it created for us in infinite variations. The tech is getting quite good, and the models we can’t access are way better than these free ones, they’re already photorealistic and coherent even with faces. 

Link to comment
4 hours ago, BabyDreams said:

This is nightmare fuel 😅

Definitely, and I didn’t post the true nightmares it created (particularly any images with faces). But it’s fascinating to see how much it already understands. And I’ve seen the big models that normal people can’t access yet; they are getting to perfection gradually. So this is a taste of what we’ll have in time.

Link to comment
2 hours ago, ihcah said:

AI can also do surprisingly decent diaper edits on demand 

I blanked out the face afterward of course, to reduce nightmares : ) Once it gets faces better this will be pretty amazing. 

image.thumb.png.fe80421ae262c589b52b8a77f6542dd2.png

It won't get better at making faces.  At least the free version of Dall-E won't.  It intentionally warps faces to prevent deepfakes or pornographic images. 

Link to comment
1 hour ago, Nimbus said:

It won't get better at making faces.  At least the free version of Dall-E won't.  It intentionally warps faces to prevent deepfakes or pornographic images. 

Well I meant “when the technology gets better” mainly. These models can be downloaded & run on your own cloud instance or hardware after all, so the more pornographic uses can be available even if they block them in their official product. 
 

That said, I hadn’t heard about them intentionally distorting faces, is that accurate? Their documentation says it’s a problem with the encoding, and really it matches other issues it has with composition of elements (if you ask for a person in a complex pose or form you’ll most often get impossible figures, multiple arms or confuse parts of legs, etc). It seems to have issues with those kinds of higher order patterns in space, and faces are particularly challenging in general. 

Link to comment
  • 2 months later...

ABDL Princess Peach

 

This shows how powerful the new tools are. It took... a long time to get this one right, tweaking the prompts over and over again and using the additional tools to cycle the image back through for a second pass when I liked it. But this is 100% AI and pretty impressive.

6.png

 

Edit: Asked for the model to make her wet, amongst other tweaks, and it worked.

image.png.805d02a1aefeb9ad8d15726486b93484.png

Edited by ihcah (see edit history)
Link to comment

These results from the stable diffusion model are impressive! Thanks for sharing - are you paying for access or are you running the algo yourself? I saw that the digital art competition at the CO state fair this year was won by someone using midijourney and a touch of photoshop, we've definitely entered a new era as far as AI generated images.

Link to comment
2 hours ago, Xeno_ said:

These results from the stable diffusion model are impressive! Thanks for sharing - are you paying for access or are you running the algo yourself? I saw that the digital art competition at the CO state fair this year was won by someone using midijourney and a touch of photoshop, we've definitely entered a new era as far as AI generated images.

I'm running the models myself, but on a paid cloud GPU (so, I have full low-level control but don't have to keep the hardware around, because few people have access to a local GPU this powerful).

I've now started experimenting with fine-tuning the image models on specific image sets. So, I fine tuned one stable diffusion on images of incontinence products (yep). Results are quite impressive with a bit of work. Note that my training images were real photos of incontinence panties (always discreet, etc) but it was able to extrapolate to different illustration styles:

Here's 2 variations of my favorite generated output from that test (prompts included housewife, kitchen, [special incontinence panty tag I trained], wet, etc):

 

887.png.2640c4351a87f9b1b658e6b6d2c4cdab.png555.png.99c5f235db4060244f27e0ada6853719.png

 

If you're wondering how variations works:  once I have an overall image that I liked, I lock in the random seed and then tweak the prompt lightly a bit. Then, I use the "inpainting" tool where I can highlight specific area of the picture, give new prompts, and have it just update that one area. So for instance I can highlight her panties and try different tags, while keeping the rest of the image intact. It takes patience but it's' really incredible what you can do.

 

 

Edited by ihcah (see edit history)
Link to comment

Perhaps I should have started a new thread now that the results are actually good rather than nightmarish. For me, I could already see the potential of those early images high up in the thread, and watching the tech evolve (and directly testing new models as they release) has been part of the fun. 

Link to comment
3 hours ago, Azazbel said:

I'm impressed by the results you have Ihcah, I guess the GPU or time needed to generate that is way too far off for our average GPUs at home?

Some higher end GPUs can handle this at home, but I certainly don't have one. Mine doesn't have the level of RAM needed... really to do customized training like this you should have at least 24GB or 32GB, on a high-end NVidia card.

But renting cloud GPUs is pretty easy these days. I pay something like $0.50 for each hour of compute.

Custom training of the model for incontinence panties took less than an hour, by the way. I'm going to do more custom training now that the initial experiment was a success. One fantastic thing is that you don't need many images to use for training -- I only put in about 5 high-quality photos of the classic white Always Discreet panties, using a really cool new algorithm for fast training by object type, and it was able to start producing images that show it understands the object very well and can represent it in many ways or mix it with other objects.

For reference, after less than 1 hour and only using about 5 reference shots of the white Always panties from different angles, it immediately could output images like this one (for a simple prompt like "rendered anime adult woman wearing <always discreet> panties, breasts, frontal view"... and this specific angle wasn't even in any of the reference shots. Truly amazing how it understands spatial things. Note the little print pattern it randomly added in the front. (I don't know why her bra is popping off like that, this is just the first image I saw it produce and was instantly amazed.)

image.png.e69ca790772d9152f2467b8b26abe965.png

If you want to try this kind of thing out, there are some guides online. The most comprehensive one is here: https://cyberes.github.io/stable-diffusion-models/

 

Link to comment

The reason for the huge leap forward in this tech is largely due to the author of Stable Diffusion choosing to release his code and his model to the world openly, so that anyone can retrain it, modify, and use it at will.

Previously, models of this power (OpenAI's CLIP etc) were carefully protected and you could only access them through a controlled service, with all NSFW content filtered. OpenAI even went further and forced "diversity" tags into its generator, so that if you put in "pilot flying a plane" they'd force your prompt to change to "pilot flying a plane, woman, black, diverse." Absolute nonsense.

With the Stable Diffusion model out there for free, on the other hand, people have already trained all kinds of custom models in just the past weeks, and you can actually make erotic art as well. It's an amazing time if you know how to use these things.

Link to comment

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.
Note: Your post will require moderator approval before it will be visible.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.

  • Recently Browsing   1 member

×
×
  • Create New...