Software Engineering Courses (SECourses)•16mo ago

Still... Seem to be faster than 2.4.0

di_en_ey•9/18/24, 10:24 AM

Dr. Gozukara, have you got any experiments on greater learning rates than 5e5 with fewer iterations? What have been your observations?

Ddi_en_ey I made a comparison between ai-toolkit vs kohya lora training. For all the same ...

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/18/24, 10:26 AM

It will outperform you are using a higher rank in kohya so it will learn more

KKallamamran WTF?! Trained a LORA on the "The Next Day" concept car from Lynk&Co yesterweek. ...

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/18/24, 10:30 AM

On how many images are you training? are you using captions? what learning rate? what network dim and alpha?

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/18/24, 10:32 AM

@Dr. Furkan Gözükara How are you mentioning what blocks to train in kohya gui? I tried this but got an error -

MMaxivi are you using with weight 1.1?

Furkan Gözükara SECourses•9/18/24, 10:34 AM

yes weight 1.1 most close

Furkan Gözükara SECourses•9/18/24, 10:34 AM

i will post an article

Ssmoke Would be strange though since all the other lora showcases it looked pretty much...

Furkan Gözükara SECourses•9/18/24, 10:35 AM

i am aobut to post an article grid

KKallamamran LR UNET was set according to @Dr. Furkan Gözükara config file (4e-6 = 0,000004),...

Furkan Gözükara SECourses•9/18/24, 10:35 AM

different workflows

Ddi_en_ey I made a comparison between ai-toolkit vs kohya lora training. For all the same ...

Furkan Gözükara SECourses•9/18/24, 10:37 AM

kohya is more professionals has more experience

KKallamamran Kohya-installation according to @Dr. Furkan Gözükara = ?

Furkan Gözükara SECourses•9/18/24, 10:37 AM

looks like xformers requires new cuda

Furkan Gözükara SECourses•9/18/24, 10:37 AM

i will just remove it

Furkan Gözükara SECourses•9/18/24, 10:40 AM

removed it

Furkan Gözükara SECourses•9/18/24, 10:40 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

KKallamamran Still... Seem to be faster than 2.4.0

Furkan Gözükara SECourses•9/18/24, 10:41 AM

yes 2.5 really faster but not as fast as linux :/

Ddi_en_ey Dr. Gozukara, have you got any experiments on greater learning rates than 5e5 wi...

Furkan Gözükara SECourses•9/18/24, 10:41 AM

it will work good on things that are not as delicate as human face

Furkan Gözükara SECourses•9/18/24, 10:41 AM

like styles

Furkan Gözükara SECourses•9/18/24, 10:41 AM

cars

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 @Dr. Furkan Gözükara How are you mentioning what blocks to train in kohya gui? I...

Furkan Gözükara SECourses•9/18/24, 10:42 AM

this is correct

Furkan Gözükara SECourses•9/18/24, 10:42 AM

re update your kohya

FFurkan Gözükara SECourses re update your kohya

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/18/24, 10:42 AM

How? I am using your installers for runpod

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 How? I am using your installers for runpod

Furkan Gözükara SECourses•9/18/24, 10:43 AM

reinstall

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/18/24, 10:43 AM

Okay thanks

Furkan Gözükara SECourses•9/18/24, 10:44 AM

single layer quality is bad

FFurkan Gözükara SECourses single layer quality is bad

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/18/24, 10:46 AM

Okay I will test others as well. Is it possible to train multiple subjects in a checkpoint fine tune without bleeding?

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 Okay I will test others as well. Is it possible to train multiple subjects in a ...

Furkan Gözükara SECourses•9/18/24, 10:51 AM

still bleeds

Furkan Gözükara SECourses•9/18/24, 10:51 AM

but each concept can be trained on each different layer

Furkan Gözükara SECourses•9/18/24, 10:51 AM

progressively

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•9/18/24, 10:52 AM

Will try that

Furkan Gözükara SECourses•9/18/24, 10:52 AM

but the quality is lower

Furkan Gözükara SECourses•9/18/24, 11:25 AM

lol even single layer bleeds

Furkan Gözükara SECourses•9/18/24, 11:25 AM

doesnt make sense

di_en_ey•9/18/24, 11:26 AM

what do you mean by bleeding?

Furkan Gözükara SECourses•9/18/24, 11:26 AM

characters get mixed

Furkan Gözükara SECourses•9/18/24, 11:26 AM

or concepts

Maxivi•9/18/24, 12:04 PM

@Dr. Furkan Gözükara yes its because flux is distilled

Maxivi•9/18/24, 12:04 PM

you cant fix the bleeding issue, even with lokr you still get bleed

MMaxivi @Dr. Furkan Gözükara yes its because flux is distilled

Furkan Gözükara SECourses•9/18/24, 12:04 PM

i trained only single layer

Furkan Gözükara SECourses•9/18/24, 12:04 PM

single layer 10 specifically

Maxivi•9/18/24, 12:04 PM

flux is cooked, sd3 8b coming out soon hehe

Furkan Gözükara SECourses•9/18/24, 12:04 PM

i guess single layer training has no value except companies to give you cheap results

Maxivi•9/18/24, 12:05 PM

single layer quality is bad

Maxivi•9/18/24, 12:05 PM

not worth it

Furkan Gözükara SECourses•9/18/24, 12:05 PM

yep

Maxivi•9/18/24, 12:06 PM

i think there is an issue with using the new torch versions @Dr. Furkan Gözükara quality seems to change and training as well

MMaxivi i think there is an issue with using the new torch versions @Dr. Furkan Gözükara...

Furkan Gözükara SECourses•9/18/24, 12:06 PM

which one?

Maxivi•9/18/24, 12:06 PM

the new dev ones

MMaxivi the new dev ones

Furkan Gözükara SECourses•9/18/24, 12:08 PM

i plan to test it actually

Furkan Gözükara SECourses•9/18/24, 12:08 PM

since multi gpu for fine tune fails :/

Furkan Gözükara SECourses•9/18/24, 12:08 PM

even on 48 gb

Still... Seem to be faster than 2.4.0

Similar Threads

Similar Threads

Similar Threads