it is exactly same as flux dev
it is exactly same as flux dev
EAh, even though it is the same model, HuggingFace expects a directory structure containing the model, encoders, and variouos JSON configuration files. I thought I might be able to copy the official HuggingFace directory and subsitute the file containing the fine tuned weights, but that wasn't straightforward. The main model contained in the HuggingFace directory was split into separate files and there was a JSON configuration file that did some type of mapping.
K
Fyou need to look for single file load if you want to use genreated checkpoints
Fmake all 8e-06
Fi had updated the config
Fafter new tests
VI need some help. I trained a flux LoRA with highest training.
So the text encoder was tuned as well.
When i place the text encoder in the SwarmUI/Models/clip path. It re-downloads with the name t5xxl_enconly.
I have this t5xxl_fp16.safetensors with me which presumably got tuned.
Everything else is fine. I am doing something wrong here with t5.
I donloaded models from this HF Repo. "OwlMaster/FLUX_LoRA_Train"
Any help would be great. I am on a cloud VM and banging my head for this.
So the text encoder was tuned as well.
When i place the text encoder in the SwarmUI/Models/clip path. It re-downloads with the name t5xxl_enconly.
I have this t5xxl_fp16.safetensors with me which presumably got tuned.
Everything else is fine. I am doing something wrong here with t5.
I donloaded models from this HF Repo. "OwlMaster/FLUX_LoRA_Train"
Any help would be great. I am on a cloud VM and banging my head for this.
VI have found some advice. Aggressive learning rate with aggresive gradient clipping. Converges way faster they say.
JLooking to finetune a model for style, and i was wondering if anyone knows anywhere I can learn about the process. I.e how to prepare dataset, what is sufficient amount of training data, what model to use etc. I'm super new to this and would really appreciate any advice or pointers. Thank you !! 
JIf it would help to provide examples of the style I'm going for, I would be happy to share
Fyou have to rename T5 to same name as swarmui uses
Fto prevent download
Fi have it
Fread carefully
JIf I'm using a synthetic dataset,will it bake the artifacts into the style?
KI'm trying to train Dreambooth flux on 29 images + regularization. Now 25 epochs have passed and the proximity to the target face is zero... Should I wait further hoping that by the 100th epoch everything will get better or is something interfering with training and I shouldn't try further?
I use kohya-ss
I use kohya-ss
Fi would say something sounds broken
Fwhich config you are using?
KI'm trying Adagrad. 20 blocks swap, which is strange for me VRAM is enough for 18 blocks, but it works slower (8 sec/step at 20 blocks and 10 at 18 blocks). Lr = 1e-5. I don't train the text encoder (I don't know if it's necessary, it's not entirely clear to me how it works when there is already T5). I use regularization and it feels like it rolls back the training...
KI also use the fp8 model as a base and the T5 is also fp8
Fis one of my configs?
KNo, I tried to put together general ideas that were found on reddit.
Ftesting grid
Fmy newest training for a client i think worked perfect
KI'm also wondering why "accelerate" isn't updated in kohya_ss scripts. As I understand, newer versions allow using fp8 weights during training. Or am I wrong?
Fwell masked training causes body disproportions
Faccelerate working fine i think
Fwell when i describe background good enough it usually works
Fduring inference
Fi still prefer just ohwx man
For woman
JIs there any way to get around that?
Fnot that i know
KHmm, without regularization with the same settings it turned out very well, from the 10th epoch the features were slightly visible, by 40 it was already very similar. Now the 90th epoch is underway, I don't see any overfitting yet.
This raises the question of why regularization complicates the process so much... Or is it all because my batch size and step accumulation are set to 1? Does it learn on 1 gradient and roll itself back when it hits regularization?
This raises the question of why regularization complicates the process so much... Or is it all because my batch size and step accumulation are set to 1? Does it learn on 1 gradient and roll itself back when it hits regularization?
KSo should de-distill flux models work better with regularization? Or should the gradient update always contain at least one target image when regularizing so as not to forget it?
KI'll probably try to experiment with Flex.1-alpha and de-distill models in fine-grained learning.
And also I wonder if fine-grained learning with block freezing will work? That is, if I, for example, train a fine model on 4 doubles and 4 singles (or any other configuration), then later, when extracting Lora of rank 256, I can remove all layers except those that were trained, getting a more compact lora, right? Or are others that I did not train also important when extracting lora?
And also I wonder if fine-grained learning with block freezing will work? That is, if I, for example, train a fine model on 4 doubles and 4 singles (or any other configuration), then later, when extracting Lora of rank 256, I can remove all layers except those that were trained, getting a more compact lora, right? Or are others that I did not train also important when extracting lora?
Fwhich guide did you watch?
Fye this is for sdxl
Fwhat you want to train?
Fok then use flux
FYouTubeSECourses
If you want to train FLUX with maximum possible quality, this is the tutorial looking for. In this comprehensive tutorial, you will learn how to install Kohya GUI and use it to fully Fine-Tune / DreamBooth FLUX model. After that how to use SwarmUI to compare generated checkpoints / models and find the very best one to generate most amazing image...
Ffor sdxl
Fwe use reg images
Fto use max reg images you need to use repeating and do 1 epoch
Fonetrainer better for sdxl
Fwatch its tutorial
FYouTubeSECourses
In this tutorial, I am going to show you how to install OneTrainer from scratch on your computer and do a Stable Diffusion SDXL (Full Fine-Tuning 10.3 GB VRAM) and SD 1.5 (Full Fine-Tuning 7GB VRAM) based models training on your computer and also do the same training on a very cheap cloud machine from MassedCompute if you don't have such compute...
