Ai image megathread

[GhostOS]

Has anyone tried putting something decent looking together with Stable Diffusion + ControlNet?

I did some testing recently, specifically with LineArt control, and it turned out quite well. They were pretty simple drawings so it wasn’t that mind-blowing, however, I tried some tickle artwork as well with more people and more complicated composition and the results are pretty good actually.

Bear in mind that this is just a quick test I’ve put together on my shitty laptop. This could be done much better by slicing the original picture and doing the smaller parts separately in higher resolutions to really get all the details right, and then merging it together in Photoshop and maybe manually fixing the rest.

View attachment 753990

Also, apologies to Augustine for borrowing his artwork for this, hope you don't mind 🙂

That is impressive, though it being Stable Diffusion it still struggles with feet (it can't do feet at all, really, unless you'd add a customly trained model I think) and extreme poses. I've used SD before (the Protogen model) through a demo space on Huggingface, but as a "smart filter" instead; putting through a rough painting in IMG2IMG with a low denoising strenght value. Still had to spend several hours of compositing, blending and overpainting in Photoshop though, but it did cut the work time down significantly. Never had much luck with Controlnet or using custom models, my computer is a pretty cheapo second hand Dell laptop with no graphics card to speak of, so local install is out of the question and I'm stuck with online demos for experimenting with. 😛

 

[odzadoro]

---

 

[RubiW]

That is impressive, though it being Stable Diffusion it still struggles with feet (it can't do feet at all, really, unless you'd add a customly trained model I think) and extreme poses. I've used SD before (the Protogen model) through a demo space on Huggingface, but as a "smart filter" instead; putting through a rough painting in IMG2IMG with a low denoising strenght value. Still had to spend several hours of compositing, blending and overpainting in Photoshop though, but it did cut the work time down significantly. Never had much luck with Controlnet or using custom models, my computer is a pretty cheapo second hand Dell laptop with no graphics card to speak of, so local install is out of the question and I'm stuck with online demos for experimenting with. 😛

Yes, unfortunately generating good looking feet is where this whole thing falls apart right now. I have some LoRA models specifically trained on feet, but it’s hard for me to run them on my current setup, so I can’t really tell whether that would solve the issue or not.

I’ve tried some photorealistic models with my own drawing and spent considerable amount of time refining the prompt to get consistent results, but it’s not ideal. Still can be improved with better reference and prompt, I’m sure.



Basically, my goal with this is not to delegate all the work to AI, but as you mentioned, to cut down the production time. I used to do photo-manipulations in GIMP like 10 years ago and one of the most annoying and time consuming parts of it was finding the right image sources. I’m okay with having to photoshop in feet and other details from a real photo, as long as rest of the image – like the poses, faces, and environment – can be quickly generated by AI.

I need a better computer first though, because just generating one 704x704 pic with ControlNet takes almost 10 minutes. And I always have to do multiple attempts before getting somewhat acceptable result.

 

[alf]

Hi great thread, tbh i don't know how these pictures are done, but some of them i like, especially the darker ones by Cavum

I’ve always been fascinated by women who love other women’s feet, especially tickling them. So I did my best to reflect this in these images of women reading magazines with women’s feet and (some) tickling in them

Yep true, this is kinky as hell just by reading 🙂

I've once sold a book, bondage and foot fetish oriented to a woman "The art of Saudelli". The cover itself showed a pretty lady being bound and of course barefoot as per the master.

Always wondered if she bought the book for bondage practice and tickling sessions 😀

 

[FFT32]

Yes, unfortunately generating good looking feet is where this whole thing falls apart right now. I have some LoRA models specifically trained on feet, but it’s hard for me to run them on my current setup, so I can’t really tell whether that would solve the issue or not.

I’ve tried some photorealistic models with my own drawing and spent considerable amount of time refining the prompt to get consistent results, but it’s not ideal. Still can be improved with better reference and prompt, I’m sure.

View attachment 754047

Basically, my goal with this is not to delegate all the work to AI, but as you mentioned, to cut down the production time. I used to do photo-manipulations in GIMP like 10 years ago and one of the most annoying and time consuming parts of it was finding the right image sources. I’m okay with having to photoshop in feet and other details from a real photo, as long as rest of the image – like the poses, faces, and environment – can be quickly generated by AI.

I need a better computer first though, because just generating one 704x704 pic with ControlNet takes almost 10 minutes. And I always have to do multiple attempts before getting somewhat acceptable result.

This is probably the first time I’ve wanted to tickle a piece of art.

 

[detritus]

Some more from fotor.com

The one called InsectTickles09.jpg (the fourth on in the list) is interesting. At first it looks as though there are hardly any bugs at all but upon closer inspection they are all inside her bikini, both the top and bottom.

 

[knight]

My own flirtations with AI tend towards the more...adult. Only drawback is it is extremely hard to get the algorithm to display actual tickling. Bes I can usually do is have the subject laughing along with the other stimulations.

 

[odzadoro]

......

 

[odzadoro]

or tummies

 

[odzadoro]

this one didn't exactly come together as planned. remember this before you use AI to write that term paper, kids

 

[surfnskate]

 

[PK Fire!]

Wonder Woman getting some love.

 

[Trololo]

 

[Trololo]

 

[Trololo]

 

[TMF Jeff]

Wow, these are excellent.

 

[michaelmichaels]

That's scarily good. AI has improved so much, even in just the last 3 months.

Are you using premium stable diffusion? I can't see how to enter an image prompt.

Here's a line art of mine I'm trying to convert.

 

[Doriansenigma]

 

[Doriansenigma]

 

[RubiW]

That's scarily good. AI has improved so much, even in just the last 3 months.

Are you using premium stable diffusion? I can't see how to enter an image prompt.

Here's a line art of mine I'm trying to convert.

Are you asking me? Not sure if anyone else here plays with SD 🙂

I use DiffusionBee because I'm on Mac and have trouble running locally anything else. If you can, you should probably go with Automatic1111.

My favourite SD 1.5 models:
- Analog Madness v6.0
- epiCRealism pureEvolution v5.0
- FML - Is that Real v0.3
- and Tickling Concept LoRA

I tried to run your drawing on my setup. Did like 10 attemps before getting something useful.



It doesn't understand what to do with those 2 lines coming from the top left corner, but other than that it kinda worked.

Here are the settings:

 

[michaelmichaels]

Awesome, thanks for taking the time to reply. Useful links, thank you. Free program. Awesome!

Handily enough I found a series of 30 stable diffusion tutorial videos on Youtube that I'll be looking through.

Link here for anybody else interested: https://www.youtube.com/@NextDiffusion

Thanks for doing the pic. I was just posting to ask if that was a good format to convert using your technique. But, thanks for the converted file. I'll post it on my DA profile and see what people make of it.
It's Gamorra from Guardians of the Galaxy. Guess the face line work could be better. 🙂

 

[RubiW]

It's Gamorra from Guardians of the Galaxy.

Oops, sorry 😀 Didn't recognize who that was. Not sure if it would do green skin without proper LoRA, though.

One more thing I've realized is that you can overlay the original artwork on top of the generated image to give it more depth. I think it looks much better + it adds in some of the details that have been lost.



I also added in a missing foot from one of the other attempts that had that part in.

 

[RubiW]

Nope, can't make her green 😀 It fucks it up too much.

I generated couple more pics, put them all into photo editor, picked the good parts from each layer, and did couple adjustments.

It's still not perfect, but it's totally possible to do good looking pictures like this even when it's something unusual 🙂

 

[GhostOS]

Nope, can't make her green 😀 It fucks it up too much.

I generated couple more pics, put them all into photo editor, picked the good parts from each layer, and did couple adjustments.

It's still not perfect, but it's totally possible to do good looking pictures like this even when it's something unusual 🙂

View attachment 754582

Really impressive, a shame my own system is too crappy to run SD locally, because a program that could turn line drawings into fully shaded art is one of my holy grails in time saving software. I'd say, it'd be pretty easy to take this image into Photoshop, select all of the skin area, and either play with the hue settings, or on an other layer, fill the section with solid green and set the layer to color, and then play around with it from there. 🙂

 

[michaelmichaels]

Wow, thanks a lot for that! Definitely impressive results.

Any uncanny valley looks down to my dodgy anatomy in the line art. This is why I'm going to continue to improve my anatomical and other art fundamentals.