TL;DR: It was trained on "pp clothes" but "pp" actively makes the output worse. Use "clothes" for the trigger word. Use "mid distance" to keep the camera from doing extreme closeups. You can specify types of clothing but your results will vary.Update May 27 2004: Turns out my trigger phrase doesn't do anything at all. I clicked the option to have a trigger phrase when training the LORA so what's up with that?Update May 26 2024: I've found that even though I trained "pp clothes" the "pp" does not contain ANYTHING from my dataset. Only the "clothes" part actually creates stains. This does mean that if I were to do another try at this LORA I know now the trigger word can just be "clothes".Model recommendations: Photo models seems to work well. Epicphotogasm Last Unicorn works well, I've had trouble with the newest version. Cyberrealistic Back To Basics works good too.The following prompt makes incredibly good images. The quality for 99% of images is very high, and you get a variety of camera angles. The dataset is all women so if you use this prompt you'll only get women.Positive Prompt: (pp clothes), side viewNegative prompt: midriff, nude, fat, (low angle), (black shirt)You can easily control the person and where they're at. Use "group photo" to have multiple people in various states of wet clothing. Suggestions for types of clothing: pants, jeans, slacks, trousers, overalls (although it only gives one style of overall), military uniform. Try others and see what works and what doesn't.Update January 19 2024I've changed captions to remove "from below" and "from behind" This seems to have reduced, but not fully eliminated, the association between "pants" and a low camera angle. You might find jeans are showing up more often when using "pants" as well.I think the bias problems; jeans, midriffs, low angle, are caused by bias of my images rather than needing better captions. My dataset has 177 pictures. Some images are from the same photoshoot, there's 63 jeans images, 40 midriffs, at least 20 low angle. The next version will take longer as I'm going to try and generate images. I don't know how many images of each type I need for the LORA to learn something. I know it's more than 2 because I have two purple dresses in the dataset and purple dresses won't get stained. I also have to be careful about introducing bias with generated images.Update January 18 2024Update to the Update: The word "pants" will bias the new Jan18 LORA to put the camera on the floor. I thought captioning prevented this kind of thing.😡 Use "low angle" in the negative prompt to somewhat mitigate this. Turns out "from below" is the wrong phrase to use, it's "low angle". In the next LORA I'll be removing those images and replacing them. I'm going to try my hand at using generated images. Then we get to find out what other bias I introduce.I've made a new version of the LORA that should make it easier to make positive and negative prompts, removes some bias, and removes some accidental associations. I captioned items that kept showing up in generation, and captioned things that were showing up where I didn't want them to show up. One example is "Military Uniform", more often than not you would get jeans. I captioned jeans, and now military uniforms no longer come with jeans.Rather than trying to caption everything I only captioned things that kept showing up like a cat begging for their next meal. Black shirts, jeans, smiles, etc. Check the caption list in Automatic1111 to see them all.I'm not able to test everything but this seems to be slightly better than the previous version. Enjoy!First postIf you find that you're getting too many black clothes add "black" to the negative prompt and increase the weight of the LORA a little bit to 1.2. This will prevent black clothes from appearing without reducing the wet stains.I'm also the creator of the "Peeing pants and other clothes." I'm creating a new page because the other is cluttered up with various LORAs in various states of working. Instead of trying to do everything this LORA only does one thing, make pee stains show up on clothing. You won't get dripping or puddles of pee. You can try using other LORAs with this one and see what you can get though.This LORA is made with a very simple caption strategy. Each image is ONLY captioned with "pp clothes", which is the activation phrase. This actually works, and works far better than anything else I've tried. After that it comes down to the correctly training it. I used this guide on the training paramaters and used their Google Colab workbook with Stable Diffusion 1.5 as the training model. <a target="_blank" rel="ugc" href="https://civarchive.com/models/22530">https://civarchive.com/models/22530</a>

Peeing Clothes (Photo)

I've removed camera directions (from behind, from below) which seems to have reduced the association between "pants" and putting the camera on the ground. "Low angle" also works better for when the camera just won't leave the ground. But an old bias has come back. Jeans will now show up where you don't want them. It's less then the first version, more than the second version. This is an image dataset issue as the majority of images are of jeans. This is not so easy to fix as I need to find are generate new images for the dataset.

Files with SHA256: c33d9cc8b2010794e3fcaa529308466fa745adae9b45972da750d5c29bba730c