AI generated fetish images

[dhf7b8g]

Links? Tutorials?

Its still quite new currently so resources aren't great. There are a few youtube videos with varying amounts of information. I tend to just skip to the training parameters part of a few different videos to copy the settings they are using. There is also this text guide which seems to be referred to quite heavily, it has reading material on loras too if you are interested in that.
I used the kohya ss gui to train, specifically a fork adapted to work better on linux.

What is it? Links? Tutorials? 😉

Its basically the opposite of an image generation model. You feed it images and it will try to figure out which traits are in the image with surprisingly good accuracy. There are a handful of choices for this.
I tend to use the WD 1.4 swin captioner as it outputs in tags rather than natural language like BLIP does. Makes prompting much easier. The model will of course only detect concepts it has been taught so chastity belt for example wasn't automatically detected which is why I added it manually.

I used this webui extension to tag the images and manually add the "chastitybelt" tag to each of the caption files. It allows you to both manually tag (which is technically better) or automatically tag using different available models.

Simply amazing! I wonder about the tags "erection", "bulge" and chastity cage" 😋

I could always give those a train 😊 Would need to collect 15-20 good quality images for each.
Although I'm not too confident in my models ability to generate guys so I might end up posting some well endowed women  😋

[Like Ra]

I might end up posting some well endowed women  😋

Even better! MUCH better!!! 😉😋

[Like Ra]

I used this webui extension to tag the images and manually add the "chastitybelt" tag to each of the caption files. It allows you to both manually tag (which is technically better) or automatically tag using different available models.

I wonder if there is a program, which can go through my huge collection, categorize them, put the tags in the files themselves or in a separate file. E.g. QNAP NASes can do that, even with a help of Google Tensor cards, but the program is very inflexible.

WD 1.4 swin captioner as it outputs in tags rather than natural language like BLIP does.

Errmmm.. Links? 😇

[dhf7b8g]

I wonder if there is a program, which can go through my huge collection, categorize them, put the tags in the files themselves or in a separate file. E.g. QNAP NASes can do that, even with a help of Google Tensor cards, but the program is very inflexible.

This is actually exactly what the extension I linked can do. It writes tags to text files and you can then filter by said tag and it will show you the images which correspond. I would imagine there are image viewers which can work with the metadata provided as its just a .txt file with the same file name as the image
Using the chastitybelt dataset as an example:
   
   

Errmmm.. Links? 😇

Both models are compatible and automatically downloaded when you select them as interrogators in the extension

Here is the link for blip: https://github.com/salesforce/BLIP

Here is the pretty barren link for WD 1.4 swin: https://huggingface.co/SmilingWolf/wd-v1...-tagger-v2
which outputs in the form of danbooru tags which you can view here: https://danbooru.donmai.us/wiki_pages/tag_groups

[Like Ra]

Excellent! The ultimate idea is to tag all images in this forum, and add a tag search capability, kinda *booru*-like.

[dhf7b8g]

I haven't trained those loras yet but I did find a few interesting ones on https://civitai.com (login required for NSFW content)
A bitchsuit and a tight rope bondage one.

The tight bondage one isnt great as it seems to just remove limbs, can't say I would use it. The bitchsuit one though...

It generates pretty badly on a realism model but its truly made for an anime model as can be seen by the previews on the page and these images:
   
   
It really likes to generate horrific feet on a realism model. Anime model on the other hand handles it much better.
   
   
   

The site has quite a lot of good content both fetish and not, Ill post a few links below

Shibari full model: https://civitai.com/models/4783/ (realism although very smooth, personally not a fan of this style)
Bondage and gag lora: https://civitai.com/models/6046/ (anime)
Latex full model https://civitai.com/models/6593/ (realism ish? Its like a realistic art style rather than photographs)
Mummification lora: https://civitai.com/models/10661 (anime but from a review it might work on realism models too)
straight jacket lora: https://civitai.com/models/10640/ (trained on anime works on both)

[Like Ra]

Amazing quality!!!

[dhf7b8g]

Lots of images in this post because I generated lots of nice stuff recently

Tried training a chastity cage lora. These AI models are notoriously bad with penis but this has to be the worst thing I've trained so far

I attached the only half decent image I was able to generate.

more successfully I trained an improved latex catsuit. The model was aeady reasonably good at generating latex before but I wanted a more realistic looking one
These two images are without my catsuit training thought I would just include them because they were nice
   
   

These images are with my catsuit training. I went for higher step, image and resolution than the female chastity belt lora so this took ~15 mins with 46 input images at 100 steps
   
   

It also does colours!!! 
   
   

and even corsets!!!
   
   

Attached some other good gens that I didn't really have a talking point for but thought I should share anyway
Will upload the lora to mega or something soon. Want to train a few more things and think it would be better to release one lora rather than three or four

[Like Ra]

Again, nothing short of amazing!

Saying that... Two points:

- the catsuits look like PVC, not latex
- also, the not perfect fit hints more on PVC

Please keep them coming! Beautiful girl, and the chastity cage looks very promising!

[dhf7b8g]

- the catsuits look like PVC, not latex
- also, the not perfect fit hints more on PVC

to be perfectly honest I dont actually know the difference I just grabbed a bunch of images tagged with latex and bodysuit. Guess they were tagged wrong.

The fit could probably be fixed with some prompting though.