I'm still playing around to determine the best settings, I'll try to keep this updated if i find something better. Download the attached sample files and see below
You need the sd3-sd3.5-flux branch to run this
I'm using windows with 24GB VRAM, try adding --fp8_base to the arguments if you have less VRAM
The UI in KohyaSS doesn't seem to support sd 3.5 properly yet, so you have to do this via command line
Make sure to adjust all paths to your copies of the files to point to the models/directories/config files
adjust Kohya_Dataset_Config.toml as needed, the directories can just be directories of the image files, you don't have to do any special folder or file naming
replace [trigger] in Kohya_Dataset_Config.toml and Kohya_Sample_Prompts.txt with your trigger word, adjust prompts as needed. Like SDXL this seems to train just fine on a single trigger word, but captions can be used
To run:
cd into your kohya directory
activate the kohya venv with .\venv\scripts\activate
run start_training.bat or just paste the command from the file into the terminal
Keep in mind Kohya samples don't always look great and can sometimes look like total garbage, but the LORA still works, test the LORAs in Comfy before making assumptions about the quality