I have very little previous experience in training anything, Flux is basically first model I've been inspired to learn. Previously I've only trained AnimateDiff Motion Loras, and built similar training nodes for it.
I can not emphasize this enough, this repository is not for raising questions related to the training itself, that would be better done to kohya's repo. Even so keep in mind my implementation may have mistakes.
The default settings aren't necessarily any good, they are just the last (out of many) I've tried and worked for my dataset.
Both these nodes and the underlaying implementation by kohya is work in progress and expected to change.
- Clone this repo into
custom_nodesfolder. - Install dependencies:
pip install -r requirements.txtor if you use the portable install, run this in ComfyUI_windows_portable -folder:
python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-FluxTrainer\requirements.txt
In addition torch version 2.4.0 or higher is highly recommended.
Example workflow for LoRA training can be found in the examples folder, it utilizes additional nodes from:
https://github.com/kijai/ComfyUI-KJNodes
And some (optional) debugging nodes from:
https://github.com/rgthree/rgthree-comfy
For LoRA training the models need to be the normal fp8 or fp16 versions, also make sure the VAE is the non-diffusers version:
https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
For full model training the fp16 version of the main model needs to be used.
- Familiar UI (obviously only if you are a Comfy user already)
- You can use same models you use for inference
- You can use same python environment, I faced no incompabilities
- You can build workflows to compare settings etc.
Currently supports LoRA training, and untested full finetune with code from kohya's scripts: https://github.com/kohya-ss/sd-scripts
