Easy SDXL LORA Training - SAME FACE, CONSISTENT CHARACTER

ThisToday I am going to show you how to make a character Lora for SDXL with Kohya & Google Colab.
I already made a tutorial how to create LORA for Stable diffusion 1.5, and the process is pretty similar, but the note we gonna use for XL Lora is a different one, and it has a lot of configurations, and it can be confusing.
So I decided to record another video, and today we gonna make a character LORA - I chose a character from The Witcher tv show, Its Yennefer of Vengerberg, and I want to create a realistic Lora that will consistently generate the character that will look like the actress that plays this role- Anya Chalotra.

0:00 Intro
0:57 Creating dataset
5:40 Set up environment
13:42 Training config
21:20 Usage
25:17 Testing LoRA

[nomad-kitchenm]

This is the most comprehensive model training tutorial I've found on YouTube. Thank you for making this video!

[keylanoslokj]

You are an AI queen!
Great content thanks.
Can you please also make a video about adding pt embedding on our colab model?

[archanameher]

I am getting an error message while running this in the training config: "/content/LoRA/meta_lat.json"

[IAGenialCrea]

Hello, I don't know if can help but Ive got two errors, I double checked and I belive I mad all the step correctly.
1-
AttributeError: module 'jax.random' has no attribute 'KeyArray'
2-
CalledProcessError: Command '['/usr/bin/python3', 'sdxl_train_network.py',
'--sample_prompts=/content/LoRA/config/sample_prompt.toml',
returned non-zero exit status 1.

[keylanoslokj]

The notebook doesn't work anymore. It doesn't connect

[PlayerGamesOtaku]

Hi, honestly this is the messiest tutorial I've ever come across. Could you redo the tutorial showing only the options that need to be used to create an XL? Also because I didn't even see the set of images to create the model where they were to be placed

[Velanteg]

This not work at all. 100 steps (10 steps x 10 epochs) returns zero result even for pose.
Current SDXL kohya requires at least 200 steps before first slight similarity appears - and its still far from any result.

[tonytitani]

I keep getting this error:
CalledProcessError: Command '['/usr/bin/python3', 'sdxl_train_network.py',
'--sample_prompts=/content/LoRA/config/sample_prompt.toml',
returned non-zero exit status 1.

[karleighrogers]

Just what I was looking for, awesome. It's starting to generate really highly realistic results as long as original pictures are in high resolution with good lighting. And only 12 pics that's crazy. You think using more pics would make it even better or not necessarily? And semi-automatic vs. automatic has an effect on final qualits result too? Ok I'll try try with my own dataset too and let you know :) Thanks again for going into each and every detail that's very helpful.