<@205854764540362752> BTW, I wanted to ask, why do you not have buckets enabled in the training conf
@Furkan Gözükara SECourses BTW, I wanted to ask, why do you not have buckets enabled in the training config?
Fi dont recommend
VWhy is that?
Fits impact being unpredictable
VI see.
Fso train at 1024x1024 first
Fthen train with bucket
Fand compare
VCan't you control it by resizing the images to be either 1024px by longest or shortest side? You still keep the max resolution to 1024x1024, but the images are then bucketed more naturally.
Fi think it is best to crop to 1024x1024
Fthen again you can just enable bucket and let kohya handle
Fi really recommend comparing results of both
Flol i have been tutorial recording for now like 5 hours
Fneed to sleep now
Frunpod part didnt finish yet
V
CHey Doc!!
I have done a first fine tune at 200 / 3000 epochs/steps and I would like to train it further.
If I use my last epoch as my starting model in Dreambooth, is there anything, any settings where I am bound to my first training?
Or can I change anything I want (LR, dataset, etc..?) for my next 200 epochs?
Can I even train it on a different Instance and class prompts to add more characters?
I have done a first fine tune at 200 / 3000 epochs/steps and I would like to train it further.
If I use my last epoch as my starting model in Dreambooth, is there anything, any settings where I am bound to my first training?
Or can I change anything I want (LR, dataset, etc..?) for my next 200 epochs?
Can I even train it on a different Instance and class prompts to add more characters?
Fyou can try adding more char but i never tried and likely to fail
Fall you need to do is decide your next number of epochs of training
Fif it was 100 so if you train 200 more it will be 300
CI used 24GB_GPU_23150MB_10.2_second_it_Tier_1.json for my first 200 epoch which has a LR of 4e-06, does it make sense to lower it as I fine tune further to avoid over fitting?
Do you look at the loss or the samples to see if you're about to overfit?.
Do you look at the loss or the samples to see if you're about to overfit?.
Flowering will reduce overfit
Fi never look loss rate
Fuseless
Fi analyze grids
Cyou mean that for each epoch you save you create all the samples in 28_Realism_Prompts.txt and compare them to keep the best one?
V@Furkan Gözükara SECourses Just finished my flux finetune, very happy with the results.
Dora left, finetune right.
Dora left, finetune right.
VGonna extract lora later and compare all 3.
Fawesome
VVI want to see (Regular) Lora, DORA, Lycoris, Finetune, & Lora (Extracted from finetune).
Fnot yet
VFconfigs are same
Fnow hours of editing waiting me 
F
SWill you cover how to half the size of the checkpoint too in this tutorial? It would be nice if we could use 12gb checkpoints instead of 24gb
V@Furkan Gözükara SECourses The extract lora from the finetune only works on the base model I used to create the finetune from. Any image created with a different model using this lora is just static.
How do I fix this?
How do I fix this?
V
VI fixed it, linked to wrong base model.
V
VThe extracted lora is actually better than the finetune!
Fthat means your model overtrained :d
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
Fdifferent model what you mean?
Fwhich different model
VI picked the wrong base model by accident.