Furkan Gözükara
patreon
OneTrainer FLUX LoRA Training Full Tutorial For Local Windows and Cloud Massed Compute
🕑 Added 2024-09-10 16:49:36 +0000 UTCPatreon exclusive posts index to find our scripts easily, Patreon scripts updates history to see which updates arrived to which scripts and amazing Patreon special generative scripts list that you can use in any of your task.
Join discord to get help, chat, discuss and also tell me your discord username to get your special rank : SECourses Discord
Please also Star, Watch and Fork our Stable Diffusion & Generative AI GitHub repository and join our Reddit subreddit and follow me on LinkedIn (my real profile)
Latest zip file : Best_Configs_v2.zip
So far I have done 8 different trainings. So the results are not conclusive yet.
Since OneTrainer still missing FP8 training, its low rank yields lower quality than Kohya
Its rank1 looked to me par with Kohya maybe better - i am gonna do more trainings and comparison
So how to use?
https://youtu.be/0t5l6CP9eBg - Full Stable Diffusion SD & XL Fine Tuning Tutorial With OneTrainer On Windows & Cloud - Zero To Hero
Above tutorial all you need. It is exactly same. I haven't tested reg images impact yet so don't use yet
EMA impact is also in test
Here Kohya best vs first 8 test of OneTrainer big grid file : test1.jpg
Check out the Grid to see yourself
Put presets into your presets folder and restart to see them
Don't forget to set your save folders, epoch, images configs etc
Currently we don't give any model path, it downloads all from Hugging Face into Hugging Face cache, no other option
As i said watch this tutorial and you can train FLUX exactly same as SDXL with preset both on Windows and on Massed Compute : https://youtu.be/0t5l6CP9eBg
Massed Compute image is getting updated to latest OneTrainer
So what are my so far findings?
NF4 works decent but adds grid like noise
OneTrainer has to add FP8 support to fix this issue - not arrived yet
Timestep Distribution : Logit works better than Sigmoid - i am yet to test others
Stochastic Rounding feels like slightly improves - no down sides
If you are not training on 40GB+ GPU, use Kohya better at the moment
Comments
Furkan Gözükara
hello. onetrainer should support atm but i havent had a chance. when prompting after the training, are you in details describing the clothing? it should fix this in flux training. also if you can find different clothing having images in training dataset it would work way better
Vlad
Hey! Are there any new changes? I have already trained on your config Kohya_FLUX_DreamBooth with my 12GB VRAM, excellent results, but I have bad dataset, out of 30 photos in dataset which 15-20 photos a person in a coat, FLUX constantly makes a person in the same coat, and does not want to change to another coat or jacket, but can change to a suit for example. I have to use inpaint or vton to make different clothes in dataset, and make a good training with my bad dataset. OneTrainer not added yet support for DreamBooth training with 12GB VRAM? I would like to train with Masked Training, maybe it would solve the problem with clothing.
Furkan Gözükara
yes i will do hopefully once FP8 gets added
楠 陈
I hope you can do more research on OneTrainer. At present, I think OneTrainer is better than Kohya.
Furkan Gözükara
no bf16_nf4 is worse. best config is rank 1. currently one trainer is bad compared to Kohya unless you have like 40 GB VRAM GPU since it doesn't have working FP8. only nf4 works and it is lower quality. i wait fp8 to do further research
楠 陈
If bf16_nf4 is better than bf16_bf16, then the best configuration should be RANK2?
Furkan Gözükara
if you make your own docker sure it can be. but i put it into a public repo atm so you dont need login etc
Antonis
is the flux base always have to be fetched from HF? Can it not be encapsulated on a docker container?
Furkan Gözükara
fixed now should work directly
Furkan Gözükara
ok you need to set access token . to save you from this step i will clone repo. give me some time and update configs
楠 陈
Cannot access gated repo for url https://huggingface.co/black-forest-labs/FLUX.1-dev/resolve/0ef5fff789c832c5c7f4e127f94c8b54bbcced44/model_index.json. Access to model black-forest-labs/FLUX.1-dev is restricted. You must have access to it and be authenticated to access it. Please log in. Traceback (most recent call last): File "E:\OneTrainer\OneTrainer\modules\ui\TrainUI.py", line 552, in __training_thread_function trainer.start() File "E:\OneTrainer\OneTrainer\modules\trainer\GenericTrainer.py", line 122, in start self.model = self.model_loader.load( File "E:\OneTrainer\OneTrainer\modules\modelLoader\FluxLoRAModelLoader.py", line 46, in load base_model_loader.load(model, model_type, model_names, weight_dtypes) File "E:\OneTrainer\OneTrainer\modules\modelLoader\flux\FluxModelLoader.py", line 261, in load raise Exception("could not load model: " + model_names.base_model) Exception: could not load model: E:/stable-diffusion-webui-forge/webui_forge_cu121_torch231/webui/models/Stable-diffusion/flux1-dev.safetensors
