Hi, it is quite common knowledge that style loras are the easiest to make and they are perfect entry point to lora creation.
Basically all You have to do is get 30-50 images, put a single trigger word and let autotagger do its job on Civit trainer.
Since recently i was sure that to get good style lora You have to keep super consistant style across the dataset and actually no.
The way i found recently and i'm alredy getting pretty good resoult with it is not to focus on the style of specific artist but rather on the theme so it makes that way of training intro some kind of borderline between Style and Concept, and on site trainer will blend in the style of images You put intro lora .
And by mixing manga panels from: Dandadan, Dragon Ball, One Piece, Berserk, Vinland Saga, Vagabond, One Punch Man, AoT, Bleach and few other mangas i got a pretty consistent and good looking blend of manga styles:
https://civitai.com/models/1472756/non-artists-styles-for-ilandnoobai-by-shrekman
The other thing that can be done in the way are loras that are focused on Aesthetic, the whole procces of making this one was 20 minutes of browing printest staring from the point of "Victorian Goth" and stuffing datafolder with pics that look cool and match that Aesthetic 0 thougs insdie of head, just going with the flow of random pintrest pics. https://civitai.com/models/1474035/victorian-goth-or-shrekman-aesthetic-loras
TLDR of how to prepare data on those loras:
Apllies to both types:
-You won't need more then 30-50 images to make it functional, more images can increase flexibility at the cost of consistency, going over 50 is worth if danboru (base dataset for IL, has a bad understanding of tag you aim for).
-Your main trriger word should be absolute nonsence i mostly go with abbreviations of the model name since if the main trriger word contains an actual functiona word that IL understands it can get You unwanted stuff.
Good example is using that style of Cherry-gig, that IL knows by itself, If You use this tag, generator will pop up random cherries in the picture even thou it understands what is Cherry-gig, it still can see a word "Cherry" and
-IL uses danboru tags, after making up Your mind what You want to train go to danboru, find simillar image and check what tags are applied for it, so You can add them as prepend Tags (optional)
-Looking at danboru tags can also help you assume how many images You actually need, if tag that You aim for does not exist on boru: 30-50 images that i consider a base for decent, flexible lore, of the concept/tag You aim for has a wild amount of entries and tag basically become meaningless like: tied hair, tied hair is oth ponytail and having a hairpinm this tag is useless, so if You want to make a hairstyle lora You might need a bigger dataset to beat intro IL what kind of tied hair You want.
The other way is You can see what tags repeat from autotagger and ad them as the optional trriger words.
-Avoid images that contain more then 2 characters, those will make lora have tendency to generate unwanted background characters.
Applies only to Aesthetic loras:
Get as big variety of styles as You can, so lora will have less inpact on the style and focus on actually Aesthetics.
Submiting Run
That's what can discourage a lot of people from training IL loras for the day i'm writing this article: on site trainer still uses IL 0.1 as base model, so You will get a terrible preview, That's the generations made with the same tags and same 2 loras.
This is WAINSFW 13:
And that's IL 0.1
The training images from training on 0.1 will look even worse, sadly i jut format the PC so i don't have any examples at hand but the bad preview in case of current Il training on side does not mean failed training.
To see what Your lora really can do You need to check it on the proper checkpoint.
Unless you straigh up train lora on using custom Checkpoint:
The only problem is that will be extra 500 buzz, but it will increase Your Lora quality, and You will have much better preview that will help You choose the best Epoch.
I'm currently training all my loras on WAINSFW to save time and make them better, lora trained on 0.1 won't be useless, generations on proper checkpoint will still bring decent resoult.
It's up to You do You want to pay extra 500 buzz to save some time.
I'm adding Mentioned models dataset as as attachment so You can check them out,.
The only seething is usually change is: get to 15 Epoch, and then Balance num repeats/batch size to get to 1800-2000 steps.
Don't bother with anything else, Your loras is ready to run and You can watch how it gonna turn out.