Can'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.

Furkan Gözükara SECourses•10/17/24, 11:40 PM

i think it is best to crop to 1024x1024

Furkan Gözükara SECourses•10/17/24, 11:40 PM

then again you can just enable bucket and let kohya handle

Furkan Gözükara SECourses•10/17/24, 11:40 PM

i really recommend comparing results of both

Furkan Gözükara SECourses•10/17/24, 11:41 PM

lol i have been tutorial recording for now like 5 hours

Furkan Gözükara SECourses•10/17/24, 11:41 PM

need to sleep now

Furkan Gözükara SECourses•10/17/24, 11:41 PM

runpod part didnt finish yet

VigilenceOP•10/17/24, 11:55 PM

@Furkan Gözükara SECourses Get some good rest.

Carlos•10/18/24, 12:29 PM

Hey 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?

CCarlos Hey Doc!! I have done a first fine tune at 200 / 3000 epochs/steps and I would l...

Furkan Gözükara SECourses•10/18/24, 12:55 PM

you can try adding more char but i never tried and likely to fail

Furkan Gözükara SECourses•10/18/24, 12:55 PM

all you need to do is decide your next number of epochs of training

Furkan Gözükara SECourses•10/18/24, 12:55 PM

if it was 100 so if you train 200 more it will be 300

Carlos•10/18/24, 2:46 PM

I 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?.

CCarlos I used 24GB_GPU_23150MB_10.2_second_it_Tier_1.json for my first 200 epoch which ...

Furkan Gözükara SECourses•10/18/24, 2:51 PM

yes

Furkan Gözükara SECourses•10/18/24, 2:51 PM

lowering will reduce overfit

Furkan Gözükara SECourses•10/18/24, 2:51 PM

i never look loss rate

Furkan Gözükara SECourses•10/18/24, 2:51 PM

useless

Furkan Gözükara SECourses•10/18/24, 2:51 PM

i analyze grids

FFurkan Gözükara SECourses i analyze grids

Carlos•10/18/24, 2:54 PM

you 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?

VigilenceOP•10/18/24, 6:33 PM

@Furkan Gözükara SECourses Just finished my flux finetune, very happy with the results.
Dora left, finetune right.

VigilenceOP•10/18/24, 6:39 PM

@Carlos Don't forget to change image repeat to 1. The default learning rate in the config was good for me, and 200 epoch is recommended.

CCarlos you mean that for each epoch you save you create all the samples in 28_Realism_...

Furkan Gözükara SECourses•10/18/24, 6:39 PM

yes

VVigilence <@205854764540362752> Just finished my flux finetune, very happy with the result...

Furkan Gözükara SECourses•10/18/24, 6:39 PM

nice

VigilenceOP•10/18/24, 6:40 PM

Gonna extract lora later and compare all 3.

VVigilence Gonna extract lora later and compare all 3.

Furkan Gözükara SECourses•10/18/24, 6:45 PM

awesome

VigilenceOP•10/18/24, 6:59 PM

@Furkan Gözükara SECourses Have you tried kohya LyCORIS yet?

VigilenceOP•10/18/24, 7:00 PM

I want to see (Regular) Lora, DORA, Lycoris, Finetune, & Lora (Extracted from finetune).

VVigilence <@205854764540362752> Have you tried kohya LyCORIS yet?

Furkan Gözükara SECourses•10/18/24, 7:08 PM

not yet

VigilenceOP•10/18/24, 7:21 PM

@Furkan Gözükara SECourses What is new in the v8 finetune config?

VVigilence <@205854764540362752> What is new in the v8 finetune config?

Furkan Gözükara SECourses•10/18/24, 7:27 PM

configs are same

Furkan Gözükara SECourses•10/18/24, 7:28 PM

now hours of editing waiting me

Furkan Gözükara SECourses•10/18/24, 7:28 PM

FFurkan Gözükara SECourses Click to see attachment

SiggiSmilez•10/18/24, 7:34 PM

Will 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

VigilenceOP•10/18/24, 8:06 PM

@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?

VigilenceOP•10/18/24, 8:08 PM

VigilenceOP•10/18/24, 8:25 PM

I fixed it, linked to wrong base model.

VigilenceOP•10/18/24, 8:30 PM

@Furkan Gözükara SECourses Extracted lora (left) vs Finetune (right).

<@205854764540362752> When can we see the newest comparison between dora, finetune, finetune lora an

Similar Threads

Similar Threads

Similar Threads