I have Compared Kohya vs OneTrainer for FLUX Dev Finetuning / DreamBooth Training
OneTrainer can train FLUX Dev with Text-Encoders unlike Kohya so I wanted to try it.
Unfortunately, the developer doesn't want to add feature to save trained Clip L or T5 XXL as safetensors or merge them into output so basically they are useless without so much extra effort.
I have tried to replicate Kohya config. The below you will see results. Sadly the quality is nothing sort of. More research has to be made and since we still don't get text-encoder training due to developer decision, I don't see any benefit of using OneTrainer for FLUX training instead of using Koha.
1st image : Kohya best config : https://www.patreon.com/posts/112099700
2nd image : One Trainer Kohya config with EMA update every 1 step
3rd image : One Trainer Kohya config with EMA update every 5 steps
4th image : One Trainer Kohya config
5th image : One Trainer Kohya config but Timestep Shift is 1 instead of 3.1582
I am guessing that Timestep Shift of OneTrainer is not same as Discrete Flow Shift of Kohya
Probably I need to work and do more test and i can improve results but i don't see any reason to do atm. If Clip Training + merging it into safetensors file was working, I was gonna pursue it
These are not cherry pick results all are from 1st test grid
Comments (14)
it's freeze in INFO Downloading onnxruntime_gpu-1.19.2-cp312-cp312-win_amd64.whl (226.4 MB) setup_common.py:189
cp312 means you have python 3.12 as default. please make python 3.10 default for kohya and make a fresh install
Have you tried fine-tuning the flux model with ai-toolkit? I see that ai-toolkit can fine-tune Flex.1-alpha (a finetune of FLUX.1-schnell). can it also fine-tune flux dev with some modifications?
I trained Flex.1-alpha with Kohya and it works. but it needs some minor parameter fine tuning since its model architecture changed .
Your comparison research keeps me glued to your Patreon.
thank you so much
Thanks for the work Dr Furkan. I used to love one trainer for sdxl. Seems a bit weird why they would implement a feature to burn time and energy training a text encoder that can never be saved 🤷
You are absolutely correct. I am quite a bit shocked by dev response
looks like Kohya is still the king
100% for FLUX
Have you experimented much with SimpleTuner?
SimpleTuner developer is a guy i would avoid. he attacked me so many times on so many platforms without any single reason :)
Thanks. So the OneTrainer currently gives worse quality? But eventually it could train the text encoder to which might help make things more accurate? Do each really need to do that yolo face segmentation thing? If you're comparing how well both of them train could that be taken out? It looks like on the 4th image in some the OneTrainer has some trouble with that - as there's some sort of box around the face in some images. Isn't one trainer supposed to be simpler too? Could it figure out the best settings based on the dataset more easily than setting them manually in Kohya (eventually - if can work properly and high enough quality)?
ye currently just stick our kohya training. box is face inpainting and model is broken so inpainting failed