Software Engineering Courses (SECourses)•17mo ago

Flux is sooo good, that it is soo tough to decide which lora is best when put side to side on a gene

Flux is sooo good, that it is soo tough to decide which lora is best when put side to side on a generator

Ddruhl Flux is sooo good, that it is soo tough to decide which lora is best when put si...

Furkan Gözükara SECourses•9/1/24, 11:51 PM

true

KKonstantinas I am trying to finetune Flux, but getting "no metadata" error. Do I need to some...

Furkan Gözükara SECourses•9/1/24, 11:51 PM

it is my this week research hopefully

Furkan Gözükara SECourses•9/1/24, 11:51 PM

fine tuning flux

Konstantinas•9/1/24, 11:52 PM

cool

I just need to use cloth lora and character lora, and combining those degrades quality by a lot. So finetuning is only solution for me now, hence researching it... But good old Lora works perfectly when not chaining

KKonstantinas cool 🙂 I just need to use cloth lora and character lora, and combining those de...

Furkan Gözükara SECourses•9/1/24, 11:53 PM

ye true

Furkan Gözükara SECourses•9/1/24, 11:53 PM

i think fine tuning may fix it as you guessed

Furkan Gözükara SECourses•9/2/24, 12:02 AM

with text encoder entered the area i am not sure about low vram ranks

Furkan Gözükara SECourses•9/2/24, 12:02 AM

i should test them

Furkan Gözükara SECourses•9/2/24, 12:04 AM

i cant decide clip skip 1 or 0 is better

Furkan Gözükara SECourses•9/2/24, 12:22 AM

majestic armor

Furkan Gözükara SECourses•9/2/24, 12:22 AM

way better than older configs

Furkan Gözükara SECourses•9/2/24, 12:42 AM

i think clip skip 0 better - learns more resemblance

Furkan Gözükara SECourses•9/2/24, 12:42 AM

Furkan Gözükara SECourses•9/2/24, 2:10 AM

generating config for every gpu

Furkan Gözükara SECourses•9/2/24, 2:10 AM

man so many options :d

Furkan Gözükara SECourses•9/2/24, 2:10 AM

Furkan Gözükara SECourses•9/2/24, 2:10 AM

i am gonna compare them

Furkan Gözükara SECourses•9/2/24, 2:50 AM

time to make new update of configs

Furkan Gözükara SECourses•9/2/24, 3:08 AM

configs updated

Furkan Gözükara SECourses•9/2/24, 3:08 AM

ranks will be also updated hopefully tomorrow

Furkan Gözükara SECourses•9/2/24, 3:08 AM

https://www.patreon.com/posts/110879657

Patreon

Kohya FLUX LoRA Training Full Tutorial For Local Windows and Cloud ...

Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon

LW24Ai•9/2/24, 4:50 AM

Who can tell me if this is normal at this rate of training?

LLW24Ai Who can tell me if this is normal at this rate of training?

it•9/2/24, 4:56 AM

You're training way too many steps. you can get a perfectly good lora in 1200-2000 steps.

it•9/2/24, 4:57 AM

unless you're training like 1000s of images

it•9/2/24, 4:57 AM

if your dataset is 15-30 images then 1200-2000 steps is enough

Iit You're training way too many steps. you can get a perfectly good lora in 1200-20...

LW24Ai•9/2/24, 4:58 AM

Except I've been practicing for 24 hours, where do I adjust the number of steps?

it•9/2/24, 4:59 AM

here

LLW24Ai Except I've been practicing for 24 hours, where do I adjust the number of steps?

it•9/2/24, 5:00 AM

too late now, you'd have to stop your training

Iit here

LW24Ai•9/2/24, 5:01 AM

it•9/2/24, 5:01 AM

how many repeats are in your img folder

it•9/2/24, 5:01 AM

Iit how many repeats are in your img folder

LW24Ai•9/2/24, 5:01 AM

it•9/2/24, 5:01 AM

just do 1

Iit here

LW24Ai•9/2/24, 5:05 AM

What happens if I shrink Epoch?

LLW24Ai What happens if I shrink Epoch?

it•9/2/24, 5:05 AM

less training steps

it•9/2/24, 5:06 AM

if you do

1 Repeat
20 Images
100 Epochs

= 2000 steps

1 Repeat
20 Images
100 Epochs

= 2000 steps

LW24Ai•9/2/24, 5:06 AM

And where do the files go after training?

it•9/2/24, 5:06 AM

Wherever you put your output folder

FFurkan Gözükara SECourses https://www.patreon.com/posts/110879657

druhlOP•9/2/24, 5:29 AM

Doc, the Lora turned out so good, even with the initial settings, that I'm going to wait till the next weekend just so you're able to release a significant update to the settings, before training again.

druhlOP•9/2/24, 5:30 AM

At this point, I believe further improvement in the input dataset will bring vast gains.

druhlOP•9/2/24, 5:32 AM

However, what I'd like to know - does Rank 1 bring more quality in the final results than say, Rank 5?

druhlOP•9/2/24, 5:33 AM

(my current understanding is that Ranks are just means to be able to train on any GPU, time varies but quality does not - am I wrong?).

druhlOP•9/2/24, 5:34 AM

@Dr. Furkan Gözükara

Iit if you do ```1 Repeat 20 Images 100 Epochs = 2000 steps ```

LW24Ai•9/2/24, 5:51 AM

Thanks

LW24Ai•9/2/24, 5:53 AM

Will Lora work with fp8 or is it only on fp16?

LLW24Ai Thanks❤️

it•9/2/24, 6:30 AM

are you using @Dr. Furkan Gözükara configs?

LLW24Ai Will Lora work with fp8 or is it only on fp16?

Neo•9/2/24, 6:43 AM

I just used his (2nd to latest) config with 20 images to ~4000 steps, with no issues. Make sure in the dataset preparation you are updating the folder repeats from 40 to 1 (saying because I missed that step). If you've already started teh training you can stop it and just rename the img sub-folder(s) to 1_ohwx man (or whatever concept it is)

Neo•9/2/24, 6:44 AM

@Laura Wolf If it has saved outputs you technically can resume the LoRA training from the latest output...

daryhead•9/2/24, 8:51 AM

I'm on runpod using a 3090 training flux with kohya and using the rank 4 file.
It's going for 3800 steps and says it's going to take 6 hours.
I'm going to stop it after about 2,000 steps, but even that is going to take a very long time.

What can I do to train in a shorter amount of time?
And is this normal for flux?
(When I used to do trainings for SDXL it was about 30 minutes to an hour)

Flux is sooo good, that it is soo tough to decide which lora is best when put side to side on a gene

Similar Threads

Similar Threads

Similar Threads