Vote here for the next Illustrious LoRA character update: https://civarchive.com/articles/9813/illustrious-lora-characters-poll
LoRA training cost Civitai Buzz. If you like my work, please consider tipping me so I can keep creating them. Reacting to my images, posting your images and using Civitai generator with my models can also help.
Description
This version is experimental. Trained based on Civitai guide to Flux.1
Include both booru tags and BLIP captions, clip skip set to 2
FAQ
Comments (11)
Can you please give me an example of one of your caption files so I can learn how to caption for flux?
for this model i used both BLIP caption and booru tags using WD 1.4 and i removed the booru character tag like hair or eye color. i don't know if this is correct, i just try different ways.
example caption:
mikasa_ackerman, long hair, multiple boys, paradis military uniform, sunlight, hair between eyes, 2girls, multiple girls, blonde hair, a group of people standing next to each other, sunbeam, day, light rays, sky, 2boys, s1, mikasa ackerman, cloud
How many images did you feed this Lora? It's very nice.
375 in pony version and 30 in flux
@andinmaro146 I see! Did you follow a certain guide for the Pony one? I'm trying to learn how to do this lol
@jbennett5901077 this is what i do for pony:
i use 50-150 (100 is very good) screenshots depend on how many quality images the character has. i don't crop the images so there are some images where the character is not solo and i try to select different angles. for tag i use the character name at start and other tags i use WD tagger and remove the character tags like hair or eye color. for paramaters i use civitai defaults and only change clip skip to 2 and increase repeats if i have less images. i usually don't go over 550 buzz per lora or 800 steps to avoid overcooking. 650-800 is good. if i have over 150 images i reduce repeats and increase epoch to keep the steps in this range.
@andinmaro146 I'll try this thanks so much!
@andinmaro146 Hey I'm going to start trying to make some and remembered his helpful advice you gave! My only question is why do you REMOVE character tags like eye color or hair? Why wouldn't you want this included? And where exactly do you put them?
@jbennett5901077 if you remove character properties it become part of the trigger word and then you can prompt just with the character name. in my samples i always put all tags but this optional
@andinmaro146 Ah I think I see. I noticed your Loras include a lot of details in the prompts though such as "black hair, black eyes, hair between eyes, long hair" Which I actually much prefer. If I wanted to Make the Lora detailed like that I would keep those tags included right? As opposed to removing them and assimilating them with the character name?
@jbennett5901077 i remove the tags in from dataset but still include in generation. if you keep the tags you have to include in generation. for my lora you can use with or without the tags both work