It is a character LoRA of Senko-san from Sewayaki Kitsune no Senko-san.
While v3A (which I personally preferred before I made v4) tends to work good only with some specific models, v4 seems to be more universal. I still couldn't get any good result with 2.5D models but maybe it's just because I didn't have any experience with them.
I had to delete tags from metadata of preview images because they contain some words civitai doesn't like. Here are couple examples of images which contain PNG-info:
https://files.catbox.moe/tfu9h5.png
https://files.catbox.moe/g4uj7n.png
https://files.catbox.moe/tvxd9d.png
My favorite model for this LoRA is Anything v5 but it should work fine with other 2D models.
I tried to train new version of the LoRA using different base models, but ended with that original NAI works the best as a base model for learning.
Description
FAQ
Comments (2)
can you pls help me with training lora of an anime character with a guide or something? i am not able to find any guide or tut on this topic
I used this guide to learn how to make LoRA:
https://rentry.org/2chAI_LoRA_Dreambooth_guide_english
(it also has some links to other guides at beginning, but I didn't check them)
So I used kohya's sd-scripts https://github.com/kohya-ss/sd-scripts and this script https://raw.githubusercontent.com/anon-1337/LoRA-scripts/main/english/train_network.ps1 which helps to setup learning settings.
My settings for v3 LoRA were:
$max_train_epochs = 20
$unet_lr = 1e-4
$text_encoder_lr = 1e-4
$scheduler = "linear"
$network_dim = 128
$train_batch_size = 10
$network_alpha = 1
I didn't use any regularization images. Total amount of image repeats per a single epochs is 7074. But just FYI learning rates and repeats count seems to be not optimal in my case.
One guy who made few good LoRA models insist that total amount of repeats (not per single epoch, but totally) should be about 2500-4000 with learning rate 1e-3 and amount of used images should be about 50-300 with setting $scheduler = "cosine".