Step 1:
Assemble 10~20 images of your desired character. make sure the character is the central focus of the image. make sure the images are between 1024 to 1600 in max resolution.
Step 2:
Put all those images into a zip file.
Step 3:
make your way over to the create button at the top of the page and hover over the down arrow.
click Train a LoRa
Step 4:
Select character
Name your lora
click next
Step 5:
Upload the zip file you made earlier
wait for it to process
Step 6:
Now you see all the images with blank text box's below them.
The first word you write in every single text box is the trigger word for your lora.
For example, if you were training a lora for naruto and all your images are only of naruto with no other humanoids or creatures in the image, write the following in every text box
naruto, solo, 1boy,
if your dataset contains images where he is wearing something other than his normal orange cloths, then also write those down such as
orange sweater,
white shirt,
black cloak, etc etc etc.
If naruto is sitting, add the tag, sitting, if they're standing, add the tag, standing,
if only their upper body is visible, add the tag, upper body,
if their whole body is visible, add the tag, full body.
If the background is white, add the tag, white background or simple background,
if you care about the style of the background, add 1 or 2 basic tags about the background, such as trees, house, building, road, forest, etc etc. and at least 1 tag about their pose, such as crouching, arms out, one hand up. etc etc etc
once you have written all relevent tags for each image in those images's text box's
click next,
step 7:
select your desired Model, I prefer pony xl for cartoon characters, but regular xl might be better for anime.
use 10~12 epochs
look at the steps counter
change the repeats counter until the steps counter displays an amount between 400 and 500.
Set the resolution to a number equal to or greater than the maximum resolution of your selected images. for example, if your largest image say, was 500 x 1100, set the resolution to 1100.
Clip skip 2 usually is a good choice
I tend to set the LR Scheduler to Constant_with_warmup.
set Network Dim to 64
set Network Alpha to 32
step 8:
click submit and wait
Step 9:
profit
