Get early access to my Anima models by following me on Hugging Face!
Over 150 models and counting!
Helpful Tags: sketch, anime coloring, black eyes
Character Tags;
Ramona, blue hair, green hair / pink hair, purple hair / green hair blonde hair , multicolored hair, gradient hair, goggles, goggles on head
Chau, black hair, side ponytail / hair over one eye, no sclera
Kim, red hair, short hair, freckles
Envy, long hair, ponytail, red hair, makeup, high ponytail, lipstick, eyeshadow
Julie, brown hair, ponytail, round eyewear
Stacey, black hair, long hair, parted bangs, thick eyebrows
Roxie, blond hair, twintails, sidelocks
demonhispterchicks, colored sclera, demon wings, blue skin, green skin, black hair, red eyes, wings, pointy ears, colored skin, demon girl, bat wings, monster girl
Scott, short hair, brown hair, thick eyebrows
Wallace, short hair, black hair
Gideon, short hair, black hair, glasses
Matthew, dark-skinned male, hair over one eye, black hair
Todd, blonde hair, short hair, tall, thick eyebrows
Neil, brown hair
Stephen, short hair, brown hair, facial hair, beard stubble, thick eyebrows
Lucas, short hair, black hair, thick eyebrows, facial hair, beard stubble, muscular, bara, large pectorals
Description
Less prone to mutant hands, a bit sharper as well.
FAQ
Comments (5)
This looks really good! I especially love how you rolled all the characters into one LoRA. That inspires me.
I have a couple questions about your training, if I may:
1. Do you crop your source images to the character, or is it okay to leave pics uncropped if the character is merely in a screenshot?
2. What are some basic parameters that you use? e.g. Learning rate, Repeats, Epochs, Network Rank, etc? I'm training a cartoony character LoRA and would appreciate pointers to reduce having to retry, since it takes so long.
1. Nope, I leave them uncropped. Not sure if thats a good or bad thing, but I find having random background characters helps the AI learn what is and isn't part of the actual character.
2. It varies on the size of the training data. For this, one, I did a unet lr of 3e-4 and text encoder lr of 6e-5. 1 repeat, for 20 epochs. I dont always use the 20th epoch, I just train 20 so I have some wiggle room in finding the best epoch.
@PopiAI
1. Cool. I uh... only took screenshots when each character was by themselves 💀
2. I'm using about 800 images for this character. With a 3060 card, this makes training take quite a long time! Perhaps 1 repeat and 8ish epochs is more my speed...
Thank you for your answers! I have never trained a LoRA before, and because of the long training times and my lack of experience, all this advice is hugely appreciated.
@AstroTibs I recommend using Civitai's Trainer, or a google collab instance of kohya_ss if you're only working with a 3060. Much faster, if you can spare the money for their respective currencies.
@PopiAI Once I get the knack for training (understanding the parameters better), I might well do that!