This workflow contains 4 groups for different use-cases. To activate one, right-click the group and click "Set group nodes to Always".
You will need to install many custom nodes, you might want to read the docs of the following repositories :
For PNG conversion install manually : https://github.com/doomy23/ComfyUI-D00MYsNodes
Group 1 : PNG Conversion
Inputs :
Path of the images to convert
Path of the outputs
Format (set to PNG)
Group 2 : Captioning
Inputs :
Path to the dataset folder containing all the PNGs (IMPORTANT : Must be PNG, not JPEG or anything else).
Trigger words and appended words to the results.
It will tag all the images one by one and create a TXT file with the same name as the PNG image.
Group 3 : Training configuration
Inputs :
Path to the dataset folder with the PNGs and TXT files.
Training prompt.
I've set the "FL Khoya Adv Config" settings to the same used by CIVITAI.com in the "Train a LORA" function of the site but feel free to add/remove network dim and network alpha, try different optimizers and so on...
Group 4 : Testing the model
Inputs :
Your checkpoint
Your new LORA model
Prompts
Size and batch
It is setup for SD15 models but can easily be changed to SDXL model loaders.
Description
SD15 LORA training works with 10 epoch or so (tested)