This is a LORA trained with Illustrious XL as the base model to generate characters in straitjackets.
Now I was in two minds about posting this. You know, I use these mostly personally, so as long as it works for me, that's all that matters. But when publishing them, I need to consider other cases. The thing is, this model doesn't behave too well with NoobAI vanilla. It works perfectly well with Cat Tower, which is the model I've been using recently.
But with NoobAI or other models, it generates ugly, 3D-looking images, often burned and fried. Why? I don't know. It was when I tested it with WAI-NSFW-Illustrious-SDXL that I decided to post it, because this one also shows anime-style images.
I actually train these LORAs with real images, because finding GOOD anime illustrations of certain fetish items like a straitjacket is really hard - many artists don't pay that much attention to detail (I'm a really crappy artist on pixiv, but attention to detail of bondage gear is something I do pay!).
But with real images, the model has a tendency to... well, generate real looking images. Cat Tower and WAI seem to be exceptions and they do generate anime looking images without having to do some heavy prompting. For other models, my testing has shown that adding many danbooru tags will eventually "persuade" the model to generate anime images, using for example a DART-based model to expand the prompt. Using artist tags usually helps get rid of the 3D-ness.
All in all, I'm providing this as is. It works for what I want to use it for, and seems to work fine with Cat Tower and WAI. For any other IL models... I can't vouch it'll work. However, using real life images to train it on has a big advantage, which is I am able to give the images a unique look - because existing art isn't that good at detail, I can do that with the LORA.
I trained the model on a dataset of around 300 images manually captioned. I've tried to train it on all sorts of angles, but your mileage may vary. I used around 25 epochs to train it on. I did a test with 15 epochs but the model was still fried in NoobAI, while not capturing some details, especially the shape of the arms under the jacket. So I settled for the 25 epoch model.
---
Instructions
The trigger word is straitjacket_IL_v1 and it will also expect canvas straitjacket in the prompt. For the neck stripe that many straitjackets have, you can use color, such as "red neck stripe" or "yellow neck stripe". Crotch strap is also another tag I used. It seems to mildly influence the generation. Probably best to put in the negative prompt if you want some seggs in your images.
Description
FAQ
Comments (8)
about the realistic issue, I wonder if it's possible to make first lora, gen image from the lora with models like cattower/wai, then train 2nd lora with image generated from 1st lora.
in theory should work right? aside from taking really long to gen and cherry picking good gen
I have considered this actually. I have several ideas in mind. One quicker way would be to tag everything with photo (medium) and realistic. Theoretically speaking the model should "differentiate" that during training and be less likely.
The only issue is they say training AI with AI is a bit... then again, at least LLMs are finetuned with chatgpt conversations all the time. Not perfect but does the work.
So yeah, I'm going to consider this. First train version 0 with the real images (I use around 200-300). I don't think I should train the model solely on AI generated images, but I could then do a second training with some of the best images from both sets, maybe 60-40 or something. It might work. I don't know I have many ideas, always trying to find one.
To be honest, my LORAs start breaking even when using WAI or Cattower if you use two or more of them together. Sometimes I may bake them too much also, train them for too long
@CulturedNiichan99 https://civitai.com/models/1223035?modelVersionId=1377943 Did my experiment on training AI with AI outputs, pretty happy how it turned out.
@user53980 nice! I've also done some experiments, but recently I'm focused again on a long running project of mine. Get a good LORA to generate handcuffs. It's the hardest thing I've found to get. You can't imagine how many experiments I've run, how many training runs, still can't get it right
@CulturedNiichan99 handcuffs real (IL) v1 by joe78, try this LoRA for handcuffs. I used it quiet a lot. And inpainted it gives excellent results, cuffs and chains.
@SD_AI_2025 oh nice, I 'm still training LORAs for handcuffs. Like... massive training, lots of versions and iterative training and over I'd say 1,000 images. hope one day I get a result I like
@SD_AI_2025 thanks, will try. I'm still trying to train my own. I'm never happy with the results lol. I must have done dozens of attempts by now
@CulturedNiichan99 Glad to see a Lora that's willing to be improved, as I'm a big user of this one so far. Trying to use it in all my generations, but it seems that currently, even using WAI, I'm confronted to semi-3d, where the image is mixed anime with 3D depth (I don't know if you know what I mean). It's giving cool results, though I'm not really a fan of for my work (most likely purely 2D). Most of the time I need to generate a pose in WAI and then inpaint the straitjacket on the subject to have what I want, but it would really really help me if you could try what you considered doing 3 month ago in your first answer to this thread. I would be gladly happy to help you test it if you want, as I really want to use this Lora to its full potential and possibly using it with other models, so if you want to, I'm your man! Cheers!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.