Software Engineering Courses (SECourses)•17mo ago

batch size 1 isnt fastest - you need to calculate effective speed

Furkan Gözükara SECoursesOP•9/1/24, 9:31 PM

but batch size 1 is best quality

FFurkan Gözükara SECourses pretty useless - you just lose quality

GomezBro•9/1/24, 9:31 PM

yup the lora was useless

GomezBro•9/1/24, 9:31 PM

bad quality

Ssergio_gonzo Dr Furkan the loras FLUX could work in any checkpoint or must be in specific one...

Furkan Gözükara SECoursesOP•9/1/24, 9:31 PM

i didnt test quantized

Furkan Gözükara SECoursesOP•9/1/24, 9:31 PM

but should work on fp8 and fp16

GomezBro•9/1/24, 9:31 PM

you had mentioned that that 200 EPOCH is for training people. What is the epoch to train a style? and why

AAfroman4peace hi all

Furkan Gözükara SECoursesOP•9/1/24, 9:32 PM

hi welcome

GGomezBro you had mentioned that that 200 EPOCH is for training people. What is the epoc...

Furkan Gözükara SECoursesOP•9/1/24, 9:33 PM

it is a good question. actually there is no hard limit. but i try to limit it 200 epoch so not very aggresive learning rate or not very low learning rate either

Furkan Gözükara SECoursesOP•9/1/24, 9:33 PM

it is like trying to balance

Furkan Gözükara SECoursesOP•9/1/24, 9:33 PM

probably you can go 1000 epoch with lower LR and get better quality but it just becomes infeasable timing

Furkan Gözükara SECoursesOP•9/1/24, 9:33 PM

so for style if you have too many images

Furkan Gözükara SECoursesOP•9/1/24, 9:33 PM

go as much as you can endure

Furkan Gözükara SECoursesOP•9/1/24, 9:34 PM

and compare every 1 checkpoint

Furkan Gözükara SECoursesOP•9/1/24, 9:34 PM

lets say you have 1000 images

Furkan Gözükara SECoursesOP•9/1/24, 9:34 PM

save every 1 epoch and compare them

FFurkan Gözükara SECourses probably you can go 1000 epoch with lower LR and get better quality but it just ...

GomezBro•9/1/24, 9:34 PM

So basically test it until you see the quality is not better but worse and find the best epoch.

FFurkan Gözükara SECourses probably you can go 1000 epoch with lower LR and get better quality but it just ...

GomezBro•9/1/24, 9:34 PM

LR?

GGomezBro So basically test it until you see the quality is not better but worse and find ...

Furkan Gözükara SECoursesOP•9/1/24, 9:34 PM

yep

GGomezBro LR?

Furkan Gözükara SECoursesOP•9/1/24, 9:35 PM

LR is learning rate

Furkan Gözükara SECoursesOP•9/1/24, 9:35 PM

use our config

Furkan Gözükara SECoursesOP•9/1/24, 9:35 PM

train unilt you see degrade in quality

GomezBro•9/1/24, 9:35 PM

you told me 10,000 images is better than 1000 right

GomezBro•9/1/24, 9:36 PM

1000 would take like 450 hrs on my a6000 with 200 epoch

GGomezBro you told me 10,000 images is better than 1000 right

Furkan Gözükara SECoursesOP•9/1/24, 9:36 PM

yes

Furkan Gözükara SECoursesOP•9/1/24, 9:36 PM

if your style very consistent

Furkan Gözükara SECoursesOP•9/1/24, 9:36 PM

even 1 epoch may be sufficient

Furkan Gözükara SECoursesOP•9/1/24, 9:36 PM

with 10000 images

GomezBro•9/1/24, 9:36 PM

https://tenor.com/view/shocked-surprised-gasp-what-cat-shock-gif-635629308990545194

Tenor

GomezBro•9/1/24, 9:37 PM

Ohhh that explains why when I did 3000 image NFT style training on SDXL a year ago it only took 9 hrs and that lora works amazing for that art style

Furkan Gözükara SECoursesOP•9/1/24, 9:38 PM

yep

AAfroman4peace @Dr. Furkan Gözükara I am here to say thanks for your tutorial and the kohya_ss ...

Furkan Gözükara SECoursesOP•9/1/24, 9:38 PM

you are welcome. also this week hopefully i am fully into fine tuning FLUX

Vvirginia @Dr. Furkan Gözükara ：（ my images from swarmUI is blurry, i do not know what is ...

Furkan Gözükara SECoursesOP•9/1/24, 9:40 PM

set CFG to 1

Furkan Gözükara SECoursesOP•9/1/24, 9:40 PM

and do more steps

JJames K Flux can run on M3 18GB Mac ?

Furkan Gözükara SECoursesOP•9/1/24, 9:40 PM

sadly i dont know did you try swarmui ?

Afroman4peace•9/1/24, 9:42 PM

I have created a lora that was based on the combined output of my sdxl model+ my flux model. after that I merged it with my base flux model. The result is very interesting. I use the resulting Flux model as "guidance" model for my SDXL model which I have trained over months. This allows me with just 4 flux steps and 15 SDXL steps with a simple upscale to create ultra detailed results.

Afroman4peace•9/1/24, 9:43 PM

I will use this output to finetune a proper flux model.. (instead of merging in the lora)

FFurkan Gözükara SECourses batch size 1 isnt fastest - you need to calculate effective speed