Use Lora weight 1.0

It won’t give good stains results with 0.8 or 0.9. It was trained to be good at 1.

It can be used to create SFW dirty clothes with the keyword (cumrag:1.0). But keep in mind it might make your grandma uncomfortable.

You will get the best results with the prompt keyword (cumrag:1.3). Attention modifier at 1.3 gives more wet cum and extreme cumrags. 1.2 gives more dry stains and dirty old clothes look. But the keyword is not mandatory.

Example: (cumrag boxers:1.3), a (black:1.1) boxers underwear left lying on the grass.

Keywords:

cumrag; cumrag boxers, cumrag briefs; cumrag clothes, cumrag hat, cumrag jacket, cumrag jockstrap, cumrag object, cumrag pants, cumrag shirt, cumrag shoes, cumrag shorts, cumrag socks.

Supporting prompts:

left lying, wearing, (sweaty:0.3), (yellowish:0.7) (whitish:0.8) (dry stains:1.2), wet stains, goo, old, stinky, gross, sticky, nasty, disgusting, (dirty:0.6), repulsive.

The base checkpoint is the original v1-5-pruned.safetensors 7.7GB. So, it should be very flexible with a variety of checkpoints. I recommend realisticVisionV51 (also the inpainting version), Reliberate_v20, objectiveReality_v20, UberRealisticPornMerge.

I DON’T recommend any checkpoint that “knows” too much about cum. VirileReality is very bad for this lora, for example, as it is probably too biased already towards man and cum. Will give too many exaggerations, distortions, and close-ups.

CFG: 5.5

Clip skip: 1. Clip skip 2 can give more consistent clothes format and composition but will mess up colors. It’s worth trying.

512px (square for clothes “left lying”, portrait for “people wearing”).

DPM++ 2S a Karras

DPM++ 2M Karras

Please use high-res fix as it will fix most of the lora problems.

Recommendation: 30 steps. Above 1200px. Denoising 0.28 for not too many changes in textures. Denoising 0.45 for better stains and textures on clothes, but if there are too many body parts on the images like legs, elbows, etc, it will get distortions, keep that in mind. The face will be great though.

Also USE ADetailer. Don’t include the lora on the positive ADetailer prompt so it can fix the face stains the Lora is prone to give. Recommend “man face”. If it is a prompt for “mask/cum in face”, don’t use ADetailer, only high-res fix. Sometimes high-res alone is better.

Since this is basically a “texture” lora, this version IS chopped of all INS and MID block layers so it will keep most of the initial composition you would get without the lora. SO, I repeat, USE IT WITH 1.0 weight.

There is an original version with all layers intact. It is GREAT for inpainting. This Lora gives great results with inpainting. Go crazy with it.

Want to have some “fun”? Install wildcards dynamic prompts extension. Here it is a good crafted script I made for testing. Paste on prompt and have fun with crazy results. Sorry if anyone gets offended by it. It was for fun and testing:

{clothes (left lying:1.35) resting displaying on the floor, (no humans:1.15), a|(standing:1.2) {handsome|ugly|fit|plumpy|muscular|short little midget|old|age 20|tough masculine|effeminate} man {from behind | | | }wearing a} (cumrag {white|dark blue|black|dark green} {clothes|one piece overalls|speedo|workgear|working suit|bikini|boxers underwear|baseball cap|coat|military helmet|bike helmet|steel helmet medieval armor|briefs underwear|long johns underwear|shoes|jacket|jockstrap underwear|jeans pants|sport shorts|socks|shoes boots|couch|pillow|shirt|cotton tank top shirt|ribbed cotton tank top shirt|face health surgical mask|cooking apron|backpack|super hero cotton costume|tie|ninja mask|bondage gimp suit}:{1.3|1.2}), on a ({gym|college|sunny beach|grass field|road|crowded night club|bedroom|couch|chair|bathroom|moon surface|amusement park|underwater|house yard|mountain|farm|crowded street|barn inside|sewage|church inside|classroom|restaurant}, background:1.25), {close-up shot|far away shot}, (sweaty:0.8), (yellowish:0.7) (whitish:0.8) (dry stains:1.2), goo, old, stinky, gross, sticky, nasty, disgusting, (dirty:0.6), repulsive <lora:cumrag:1>

Current problems with the Lora:

• face distortions.

• sometimes cum on bare chest.

• sometimes concept bleeds to the background.

• too much ribbed cotton textures, use “(ribbed cotton:1.3)” on negative prompt.

• some bad close-ups when “wearing” the underwear, normally not a problem.

• sometimes too much cum, not realistic, too many stains (? maybe it is as intended).

• hats need more flexibility, too much biased towards caps.

• jockstraps are really bad, even though I had a whole folder/concept for it, it probably need more full body data, people wearing it.

• Need more towels on dataset. Maybe in the next version.

• Not very flexible with other “texture” keywords like flannel, patterns, stripes, etc. Avoid them but it might be good to experiment with cloth materials. Ribbed cotton might be ok as there was a lot of it in the data set and it was captioned most of the time. It’s actually biased towards it, as I mentioned before.

• I have not tried too many results with women, but I can’t imagine why it wouldn’t work. All dataset was on men. Also, no data on anime, comic, drawing, but from my testing it does it well.

Some settings: Trained 768 res. More than 500 pictures. 10 steps each image. 12 different folders/concepts (sharing the same “cumrag” idea). AdamW.

Some history:

This is actually my first lora. But I’ve been redoing it for months now. This is the 13 version. It started from the complete inability of SD to create ANY dirty clothes. I’ve wanted to create something that SD didn’t really know, so I started this “project” searching for it. Eventually I got to the dirty corners of the internet and “cumrag” was the obvious choice. It proved to be a great starting point to get to know all the Loras learning parameters, data preparations, etc. This was not easy (although my first version was already pretty good) because all tutorials are about characters, styles, or objects. This is none. It’s kind of a concept, kind of an object (clothes), probably leans more towards a “texture”. And it can be applied to many different things, colors, all types of clothes, and different objects.

As time has passed, I’ve learned some keywords that would help normal SD to do dirt, but basically any result without a lora is too clean and bad. SDXL is way better at this and can create stained old clothes way easier with simple prompts. I do intend to retrain this on SDXL soon.

Khoya training can be read from the file. I was not very happy with Prodigy and other adaptative optimizers. AdamW gives more constant results. My experiment with LyCORIS wasn’t good either. Caption does improve results, but not as much as people would think. It increases flexibility a lot. For example, my first version had a wrong settings so all my “.txt” files with captions were ignored (pay attention to the caption options on khoya). But it was pretty good, maybe the best at some clothes. But it could not change the background at all. Like for example “left laying on grass”, even with heavy weighted prompts.

Also, for anyone doing anything similar, please have a look at lora block layers! It can give good results to train with different weights, but with my experience, even after deciding what layers is the most important, using the extension “Supermerger” to stripe the layers after training gave me better results than starting the training with different weights. But I did not heavily tested this.

I do intend to keep developing this, not because I’m that interested in the concept, but I think of it as a good challenge. I hope to fix its defects by using regularization images of man and woman with perfect faces and more data images with whole body. I will feed the training with good results from past versions too.

I have no idea how people will react to this. But hopefully you can leave some testing of your own and some comments. Any idea is appreciated. Thank you.