AI generated fetish images

[Like Ra]

libcudart.so.11.0: cannot open shared object file

Isn't it related to NVIDIA's CUDA developer library? Some years ago, only old CUDA versions were supported by TF.

I install Tensorflow for my AMD card

Does TF support AMD (via OpenCL) nowadays?

[Like Ra]

Under Windows you can use DirectML for AMD GPUs: https://community.amd.com/t5/radeon-pro-...a-p/488595

For Linux, TF + PyTorch is the answer.

I'm on Linux. I did manage to get everything configured and working on my AMD card, but as I said, 8 gig of VRAM was not enough so I couldn't generate any images. I used ROCm to have my gpy working with Tensorflow and Pytorch.

This Python stuff allows me to get some images out of it but it's slow, a set of 8 images takes almost an hour on my six cores and 16GB of ram. If I do something else while it's running I run out of ram and it crashes.

On my PC one image takes about 10 secs. I tried some different words. "bald woman " gives me men looking woman. Interestingly it knows some what "Zentai" ...

krin

On my PC one image takes about 10 secs. I tried some different words. "bald woman " gives me men looking woman. Interestingly it knows some what "Zentai" ...

"Body suit" or "Full body suit" works well, "Leotard" will also be recognized, although sometimes with gloves fused to the skin. As for materials - spandex and latex both yield good results. So shiny!

           


Sometimes it even spits out images with great bondage posture, but clean restraints are just a no-go.

   

I feel I should move on from latex suits, but it's hard letting go.

Such a shame the arms in the second image are messed up beyond fixing, the fusion of yellow and transparent latex on the legs looks intriguing.

[Like Ra]

I still wonder about the prompts.

I still wonder about the prompts.

All are basically iterations/variations of

Code:

beautiful Sarah Shahi, wearing skintight yellow latex costume, see-through, translucent, transparent, digital painting, artstation, concept art, smooth, sharp focus, illustration, art by artgerm and donato giancola and Joseph Christian Leyendecker, Ross Tran, WLOP -s75 -W512 -H768 -C14.0 -Ak_euler_a -G0.8 -U 2.0

Adding the Seed -S4106275009 for istance should give you an exact copy of image #2.

The translucency/transparency will get ignored it appears, sometimes the AI just adds it on a whim.

I find that euler_a gives the most aesthetically pleasing results. Unfortunately, the sample mode is not deterministic, so even the slightest variation of the prompt (even the number of steps) will give you drastically different outcomes. Other samplers are more lenient if you want to keep the general appearance and just want to change a tiny detail.

What works really well is taking two very similar images and use some editing to fuse them together.
The last one for instance initially had a deformed hand, so I ran variations until I got one with the glove.
Since the poses are almost identical, you can just go and overlay them and mask out the parts you don't want.
I also went and masked in the shinier torso.

I'm still experimenting with the various parameters, I'll post when I have a summary of my findings ready.

For now:

- specify a name of a celebrity, otherwise you'll get weird results. You'll still get weird faces but less 😁
- more steps doesn't always mean better results, I got some good results with only 7 steps.
- you can use a seed, I generate lots of low step results for various seeds and then use the seed of one that turned out interesting to generate versions with more steps. I only need to do this because I'm running on cpu, 50 steps takes over an hour (as apposed to 10 seconds mentioned above)

more tips: https://www.howtogeek.com/833169/how-to-...on-prompt/

prompt should follow this template for better results: "A [type of picture] of a [main subject, mostly composed of adjectives and nouns -avoid verbs-], [style cues]*”

[Like Ra]

Still waiting for low/normal prices for RX 6700XT... 12GB should be enough, I hope...