Cum On Cloth Lora
Description
Negative prompt if there is too much cum on the head should be: cum on head
FAQ
Comments (14)
Nice!
That's awesome! I've just released my cum on clothes Lora this week!! I'm testing how well it plays with yours! Mine is more focused on stains, males and "cumrags". I think they complement each other quite well!!
From my testing, yours at 1.0 will not make clothes left lying on the floor, of course, as it was probably not trained for that. It will play quite well with mine at 0.5-0.8. Helps to make more extreme cumstained clothes.
Mine will help yours at 0.1-1 but you might get more face distortions at high weights (and too much face cum).
Not composition distortions or overcooking signs from both our Loras, which is great!
I'll probably use you lora with mine all the time now =P, it works great with the new leco detail slider from Ostris jaretburkett651
@diogod I'm glad that you liked it. I will test your Lora ;)
This is absolutely phenomenal. It really goes well with the downblouse Lora.
I've been trying to train Loras myself, and although I have been able to capture concepts like this, I have not been able to do as good a job as you did.
Do you mind sharing some details of how you trained it. How many training images, what were in the training images, network ranks, regularization?
Want to learn your technique!
Hi. Thanks for your kind words. I didn't find a magic bullet for training Lora. But for now, I can say that it's better to use 20-30 steps and not 40 for each folder (example: 20_imagename). I'm using more than 30 images per training. Train batch size 2-4 (for style). I think the main idea is after using WD14 for the caption remove all unnecessary (I mean very small details) tags. If it's background or leash or some other staff it should be added as a tag. I make 10-20 epochs and check them with different prompts with 'Prompt S/R'. Dunno what to add. Nothing special I think. Hope it will help
@AiTimeIsNow Thank you for taking the time. I will try this. May I ask what Network rank and alpha you're using?
@rollpollution193 ok. I don't use 8 rank because it's too low for the style (IMHO). 64 - I use the most if I have some style with different poses or some strange things but not too much with 1 alpha. When it is 128 you can put a lot in this Lora and use 8 alpha (it's just multiply scales if I'm not wrong). To test all this stuff better use LR Scheduler - const with the same Seed like 123. I don't write in English often so it can be hard to understand what I'm trying to tell.
@AiTimeIsNow thank you!
I want to make a request for this in SDXL!
sorry but for this my PC is too weak for that:) even in 4080, it is so slow that I can't make any normal lora for sdxl I don't know why.
@AiTimeIsNow Do you mind uploading the training data? I'll train it in SDXL and share the model with you!
@rollpollution193 sorry but I've lost it and it's sad. Maybe I will create remake if you want
@AiTimeIsNow Oh no! haha. If you can, please. Your model is one of the best I've used.
It works with other Lora so well. I want to see what you did :D