Prepare training data

example training data:

https://drive.google.com/file/d/1PQ_I1mqLPJ5QxqqKWtXS3UpJ292-jmkQ/view?usp=sharing

download and unzip

for example your unzip location would be:

D:/samples/

and inside it :

D:/samples/1_fernsoso_girl/001.png & 001.txt .. .. ..

Install Git

https://git-scm.com/downloads/win

Download flux.1-dev

https://huggingface.co/black-forest-labs/FLUX.1-dev

Download kohya ss

Then we need to download kohya ss to our machine.

go to this github repo and clone it out.

https://github.com/bmaltais/kohya_ss

for example your local disk location is :

D:/git

then:

cd D:/git

git clone https://github.com/bmaltais/kohya_ss.git

Install Python

We also need to have a python installation in our machine as well. Kohya ss recommends Python 3.10.11 so we’re gonna use that.

https://www.python.org/downloads/release/python-31011/

Download Windows installer (64-bit) and install it to your machine

for example your python installation location is :

D:/Python-3-10-11

Create Python virtual environment

Go to your local git repo disk location:

cd D:/git/kohya-ss

since Flux training codes aren’t merged into main branch yet, we need to checkout sd3-flux.1 branch

git checkout sd3-flux.1

then we need to create a python environment to install the required library

mkdir venv

cd venv

D:/Python-3-10-11/python.exe -m venv flux

Install the required libraries

after the virtual environment is created, go back to base dir:

cd D:/git/kohya-ss

then activate the virtual environment:

./venv/flux/Scripts/Activate

After that we can start installing the python requirements

pip install -r requirements.txt

pip install torch==2.4.1+cu124 torchvision==0.19.1+cu124 xformers==0.0.28.post1 --index-url https://download.pytorch.org/whl/cu124

pip install -r requirements_windows.txt

Launch kohya web ui

Now we’re ready to launch kohya web gui

python kohya_gui.py

if you see this then it means you can launch kohya web gui successfully

in your web browser, go to the url displayed above. Usually it will be

http://127.0.0.1:7860 or http://localhost:7860/

go to LoRa and Training tab

Next we need to fill in the Flux.1 related values, i.e : VAE Path, CLIP-L Path , and T5-XXL Path

The first two should be available from the flux1-dev download directory

As for T5-XXL Path, our flux1-dev text_encoder_2 dir contains models that are split into two files, but this field only accepts single file so we need to find a single file safetensors for it.

I managed to find one below. Download the t5xxl_fp16.safetensors and put the file path to the field there

https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main

Now we are ready to train, just hit “Start Training “ button.

Training finished

After the training is finished, go to the output folder you specified above, and the LoRa safetensors would be there. In the cases of settings above, the LoRa file will be:

D:/training_output/fern.safetensors

How to apply LoRa in diffusers