Values to train a style.

[TetraOrb]

Hello! I tried to train a style, but the final result looks nothing like the style I want. So I'd like some tips on values. I use a set of 288 pictures. When I use all of them as instance pictures without I get a problem with faces, that do not stylize at less than 7000 steps.

So I tried to divide them into two parts - 50 instance images and 238 concept images. I used 10000 steps, 350 text encoder and 2000 concept encoder, but the result looks nothing like the style I wanted.

Here is the comparison of what I want and what SD draws after training. I'd be happy to get help to adjust the values to get the style closer to mine.)

[TheLastBen]

10.000 steps is too much, reduce the instance images to 15-20 and the steps to 3000, keep the text_encoder to 350

[nawnie]

Oh no I was wanting his original set :P style is nice, but some of these pictures contain the same subject, I feel a little change up in source data and we can get a huge diffrence.

[TetraOrb]

here is the model https://drive.google.com/file/d/1-3utQxN5_R6uD5oDSmeiUmCqaCYrVL5u/view?usp=sharing but don't expect much from it as the instance images are inconsistent and low quality.
identifier : bbvvtt

Thank you! Let me ckeck it out!

[TetraOrb]

Oh no I was wanting his original set :P style is nice, but some of these pictures contain the same subject, I feel a little change up in source data and we can get a huge diffrence.

Here are my pics: https://drive.google.com/drive/folders/1DzZ022a_fZqjcjsUjDSpGPUqg6JVBO4j?usp=share_link. Each folder has a folder with cropped and uncropped images inside)

[TetraOrb]

This could be an application for aesthetic gradients, works better for training an art style

https://arxiv.org/abs/2209.12330

Sounds interesting! Thank you!

[nawnie]

i can try again with the full set if you want, if not the upscaled set i made from what you
shared earlier is in my link underneath

[nawnie]

enjoy,
keyword is petrova, ive been using the prompt in xxx,the style of petrova 20 steps cfg 7 eurlerA

... imma ust say no adjustments prompt and go bud

https://drive.google.com/drive/folders/1GC81k3LI9RwHMcRLOt513MPEeIahGgU2?usp=sharing

[nawnie]

hmm looks blurrier big one sec

[nawnie]

if im being honest, it could use another round of training, i wanted to capitalize on style training, so my suggestion is to take images and run it at 700+ resolution for another 1000 steps saving every 200

[TheLastBen]

the instance images are of low quality so the model is mimicking that