Click on the images for HD previews
Tags for Domina
domina112, 1girl, gun, jewelry, black hair, dark-skinned female, jacket on shoulders, blonde hair, high heels, multicolored hair, belt, pencil skirt, nail polish, pantyhose, mechanical arms, white skirt, extra arms, lips, cleavage, clothing cutout, two-tone hair, blue nails, large breasts, cleavage cutout, mole on breast, gold trim jacket, gold bracelet, gold bikini, black top,
Bless the OW devs for cooking up these great characters 😌
Description
FAQ
Comments (5)
What base model are you using for your LoRA? Did you merge it yourself? How did you achieve such a strong personal style? How many images and epochs are in your training dataset? I've recently been wanting to learn LoRA training.
Normal base model. No merging. For dataset you want GOOD images not a lot of images. 40-60 is where I like to get to. I edit all the images to remove elements that would not be good for training bright lights, shiny surfaces stuff like that. I would look up a video on how diffusion models work on a technical level for insight on how to spot what to look for. My settings for training are mostly the default ones that came with onetrainer for sdxl. Number of steps in training are different depending on the size of the dataset (you need to do the calculation each time) but with the dataset sizes that I like to use normally around a few thousand. Always make sure the images is at least 1000p x 1000p. The most important part is the dataset. I spend way too much time on every one I put together.
@K112 Thank You for info, I was about to ask same question.
I just wonder if You have experience with https://github.com/ostris/ai-toolkit is it as good as onetrainer? It would be nice to have someone experienced, like Yourself, to voice opinion regarding that.
As for removal of elements I tried to do that for manga, that had colours, did used different colour, but close enough, to overpaint i.e. text bubbles and... lora did picked it up :) and I ended up with "patches of clouds" in my test images ;/.
Could You recommend some quality content about "diffusion models work on a technical level".
Thank you in advance.
PS. what about training on base SDXL and than using it in IL/Pony? I do assume that training on different base model will result in different LORA weighs and outcome, but as IL/Pony is SDXL based some of these do exist in IL/Pony. Anyway what is You experience, if any, with such merging models?
@K112 Thank you very much for your answer, thank you ❤️
@N0n4m3 Part of this video explains it well but you could find more in depth examples https://youtu.be/NsM7nqvDNJI?si=t4qVj2i8Wf0AHU3M&t=825 . I do not have any experience with that tool if it works for you and you like it keep using it. Don't over paint text bubbles unless they can be removed perfectly like a blue sky background for example. If the image is tagged with text bubbles it is fine to use that for training just make sure the image is tagged. Removing the text from the bubble and just leaving it blank is what I do. If you are making an IL lora train on base IL not sdxl.