Description
Work in progress flat chastity lora trained on base 1.5. Works best on realistic checkpoints, preferably with no knowledge of genitalia (though nsfw checkpoints can still work). Recommend realistic vision. Less likely to work well on anime/stylized checkpoints and on clip skip 2.
Highly recommend using hires. fix 4x_NMKD-Superscale-SP_178000_G at 0.3-0.5 denoising strength to improve cage details.
Prompting for cage
{optional color} flat chastity cage
{optional color} flat chastity cage with {optional color} {optional material (string)} belt
"flat chastity cage with black belt" was the most common tag in the training data, so it should generally be the most reliable.
Description
Basic workable version with issues with body proportions and crotch-focus
Warnings and Tips
The training data largely did not have faces and were often zoomed in crotches, so proportions can be off and faces are hard to get to show up at full strength. Putting "close-up, cropped, head out of frame" in negatives can help.
You can also try generating at 0.75 strength and inpainting afterwards to preserve proportions. If inpainting, try using "only masked" and add "close-up, cropped, crotch" for better detail.
If generated cage has low detail, hiresfix can help.
FAQ
Comments (9)
So, first try.
The problems that I found (in one-click-gen)
1) sustainable view from below
2)sustainable banded-grainy texture in the background
3)sustainable "piss-filter"
(in inpaint)
4) not very good with scaling relative to the rest of the image
5)not very good with color and shadow match
I'll share my results as soon as I get them good enough. So far, this lora gives the most realistic cages, of those that I tried, I was able to get flat cages on other models, but every time it was pure randomness. Your LoRA gives the best surface and clear shape, but it takes a lot of attempts to generate this shape and surface in the right place and position
Yeah, those are the issues I have too. Unfortunately, my data set was mostly middling quality to lowres crotch closeup selfies and torso shots with the head out of frame from below. I did tag some images with "from below" so you could try putting that in the negative prompt. If you know of any good sources of high quality full body shots of guys with chastity cages, let me know!
For the grainy texture, that's something I haven't consistently tagged for in some of the lower res source images. You can try putting "lowres" in the negative prompt however.
Not sure what you mean by "piss-filter", but if you mean the lighting, yeah, I haven't tagged for that in training, so unfortunately, that's probably gonna be an issue.
I'm still working on the lora (hence the 0.1 version) and gathering more and hopefully better images for training, but it's slow going. Flat chastity cages seem to be so much less popular than the regular black plastic ones.
what checkpoint are you using by the way?
@out_of_username_ideas by "piss-filter" I meant tight cyan/yellow color tone, and burned contrast throughout the image. In one-click-gen, it decreases at a value 0.7, but does not go away completely while this LoRA is in the promt.
I haven't tried adding "lowres" to negatives, I'm too used to using textual inversions. I'll definitely try this. And, I added "from below" to negative, but its need high values to work.
Now I have got good results with Perfect World, semi-realistick hentai model, with knowledge of genitalia. I also tried AbsoluteReality, I don't know how much does it know about genitalia, I usually uses it for gen DnD characters, and as "help-layer" with any NSFW LoRAs. I will continue to try other checkpoints.
So, I don't know good source for collecting database. Anyway it takes a lot of time. (may be i can help)
@lesard Thanks for the feedback! I'm guessing the "piss filter" is from some weird lighting from some of my training images. I'll check if I can tag that or remove those images the next time I train.
For negative prompts, you can also try "close-up, cropped" and maybe "crotch", since I added these tags to all of my crotch close-up images.
For checkpoints, the two checkpoints you listed seem to recommend clip skip 2. This lora should generally perform better on checkpoints trained on clip skip 1, since it was trained on the base 1.5 model at clip skip 1.
Personally, I only use realisticVision and Virile Reality myself.
I've found a way to generate good chastity with https://civitai.com/models/132388/penis-pussy-adetailer-model. I use lora and (flat chastity cage:1.4) in adetailer prompts with high denoising strength (around 0.8) leaving main prompts without anything of it.
Man, you are best.
It's first Lora, which generates normal chastity cages.
I test it on my custom checkpoint (merge epiCRealism and PornMaster) and received good result on the first try.
But I'll share my results, when can do it perfect...
I finds this problems:
1. Size, it's makes too big cage (on weight 0.65), if less weight, cage blurring.
2. This lora spoils the skin texture. It's makes pubic hair, and if add tag "shaved", that doing many problems skin, for example shaved marks, red and black dots, but not shaved.
Glad to see you managed to get some use out of the lora!
For the proportionality, hopefully, I'll be able to solve that on future releases. I may combine the lora with other chastity cage types though since flat cage pics are hard to find and often close-up crotch pics however.
For pubic hair, I tagged the hairier pics with "hairy {ethnicity} man", so you can try putting "hairy" in the negative prompt.
I usually use "pubic hair" in a negative, and it always works better than just "shaved" in a positive . But I have not yet used this LoRa on realistic checkpoints, only on the most “glamorous” ones