Flux version added
for that late 80's early 90's anime aesthetic
Trained on Satoshi Urushihara artwork and OVA stills.
Description
this is a huge 384 dim LoRA, experimenting with adaptive learning rate
note I tried resizing this LoRA to lower dims using the kohya_ss tools, but the quality results were too subjective (some images degrade, others are improved)
training settings
3137 training images
512,512 training resolution
16 epochs, 6,384 steps, 8 batches
network dim & alpha 384
DAdaptAdam optimizer, constant scheduler
learning rate 1.0
bucketing & random crop
useful optional tokens
urushisato tagged on all images for general style boost
ova tagged on all screencaps from animation
background tagged on any image that has no human subject
FAQ
Comments (3)
Amazing work! I also wanted to make a lora based on Satoshi Urushihara, but after seeing your work and the number of arts, I immediately changed my mind, I just physically achieve such a variety and use so many arts, thanks for your really great Lora! (👍 ͡❛ ω ͡❛)👍
Hi, your math on step count is wrong.
16epoch * 3137 TotalImagesWithRepeats / 8 batches= 6384 Steps total (thats 399 Steps per epoch). Batches reduce the step count, since you train more images at the same time.
漆原智志啊，太强了