Is there any way to get around that?
Is 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...
T@Furkan Gözükara SECourses IS there any guidance / video from you to teach us how to train Lora for Wan 2.1 ?
�Guys, could a person with tattoo be trained and the Tattoos be reproduced correctly? Would sdxl be able to do that too? Fine tuning @Furkan Gözükara SECourses
TI can tell u it's very hard for SDXL, maybe possible for Flux
Fhopefully will make but i still couldnt have time yet 
Fbut flux is doing i trained dwayne johnson
TCan your video to educate us on Wan 2.1 Lora involving both Local Computer Training and Massed Compute Training ? For now, can u tell me if I can train Wan 2.1 on kohya ss for Lora ?
Fyes will do hopefully
Fsorry for delay so many stuff 
Fwan 2.1 full quality lora is my priority task
�@TrumpYouAll I'm curious if including photos zoomed in on tattoos with correct tagging would improve details and consistency 
Tthanks Dr, I would liek to use Massed Compute as well as local computer to train Wan 2.1 Lora. Thanks Dr ! @Furkan Gözükara SECourses +
TSorry, no knowledge on tattoos but can't be done with SDXL
CHow should you determine which base model to train your dream booth/lora on? Is there a good comparison sheet somewhere
TThe best model to start is always the 'base' model , not any 'trained' model
Fyes will cover both hopefully
Fflux base
Fbest
CIf it's objectively best why are so many people still training on other models nowadays?
Ftesting :d
Fi tested these for example : https://medium.com/@furkangozukara/thoroughly-experimented-with-fine-tuning-dreambooth-training-of-flux-dev-de-distill-pixelwave-13fab8c625bf
Medium
Thoroughly experimented with Fine-Tuning / DreamBooth training of Flux-dev-de-distill, PixelWave v03, Verus Vision and base FLUX Dev model…
Coh cool, thanks!!
NI did one training on Flux DeDistilled, because negative prompts were important to me, and I am very satisfied with result. I didn't see any objective benefits from base model over the de-distilled one - and negative prompt is a huge benefit on the other hand.
that said, I guess it all depends on your personal goal and needs.
that said, I guess it all depends on your personal goal and needs.
AYes a little, but in the end it will not work
AI have tried extensively. Same for piercings
AI had some small luck with a piercing https://civitai.com/user/Aikage you can see my Sanne character has black ear clips but they're not always consistent.
AYou are correct the best way I have found is to take a lot of pictures of JUST the tattoo and tag it something, then take pictures of the character with the tattoo and without and when the tattoo is present tag it with the tattoo tag which will link it to the close up pictures. Still it mostly just gets the rough shape correct
HBased off the example 256 img training log for the 48GB-batch size 1 config, should I expect to train until ~50K steps before seeing decent results?
F50k probably overkill
Fi had best checkpoint at 70 epoch
F17920 steps
H
what epoch was your best checkpoint for the batch 7 config? I have 252 imgs so I’ve been training until 111 epochs to reach 4K steps but never went higher than that.Fi dont remember but probably similar
Fyou know epoch doesnt change but step count changes