Software Engineering Courses (SECourses)•17mo ago

starting a training with TE

Vs12•8/27/24, 6:32 PM

i've been using it this whole time

Furkan Gözükara SECourses•8/27/24, 6:32 PM

https://imgsli.com/MjkxNzAw

VVs12 i thought it was common to use shift value 1

Furkan Gözükara SECourses•8/27/24, 6:32 PM

i was using default value 3

Furkan Gözükara SECourses•8/27/24, 6:33 PM

also this is shift value sampling

Ddsienra starting a training with TE

dsienraOP•8/27/24, 6:33 PM

let's see if still bleeds......

VVs12 i thought it was common to use shift value 1

Furkan Gözükara SECourses•8/27/24, 6:33 PM

you mean discrete_flow_shift right?

Furkan Gözükara SECourses•8/27/24, 6:35 PM

Furkan Gözükara SECourses•8/27/24, 6:35 PM

colors better

Furkan Gözükara SECourses•8/27/24, 6:36 PM

armor better too

FFurkan Gözükara SECourses https://imgsli.com/MjkxNzAw

Vs12•8/27/24, 6:36 PM

Overfitted still

Furkan Gözükara SECourses•8/27/24, 6:36 PM

Furkan Gözükara SECourses•8/27/24, 6:36 PM

although likeliness a bit reduced

dsienraOP•8/27/24, 6:50 PM

runs slower on WSL with TE

Ddsienra runs slower on WSL with TE

Maxivi•8/27/24, 6:50 PM

did you upgrade to torch 2.4.0?

dsienraOP•8/27/24, 6:51 PM

yes I did 2 days ago, I was having 3.5 to 4 it/s

dsienraOP•8/27/24, 6:52 PM

now I enabled the text encoder and it runs much slower

Maxivi•8/27/24, 6:52 PM

you training 1024?

dsienraOP•8/27/24, 6:52 PM

yes

dsienraOP•8/27/24, 6:52 PM

something is on cpu

dsienraOP•8/27/24, 6:53 PM

I think is T5 that is not training, just CLIP_L

dsienraOP•8/27/24, 6:53 PM

but really idk

dsienraOP•8/27/24, 6:54 PM

may be the performance can be improved

VVs12 Succesfully trained a model with clip l text encoder, results are better now and...

dsienraOP•8/27/24, 6:56 PM

what about the speed training the text encoder? is running slower for me

Furkan Gözükara SECourses•8/27/24, 6:59 PM

new findings shared : https://www.patreon.com/posts/110293257

Patreon

Kohya FLUX LoRA and Fine Tuning Training Full Tutorial For Local Wi...

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

Furkan Gözükara SECourses•8/27/24, 7:05 PM

time shift sampler comparison

Furkan Gözükara SECourses•8/27/24, 7:05 PM

https://imgsli.com/MjkxNzAy

Furkan Gözükara SECourses•8/27/24, 7:05 PM

configs added to folder

smoke•8/27/24, 7:06 PM

@Dr. Furkan Gözükara I am having a OOM error in Forge, on Massed Compute, do you have any idea?

smoke•8/27/24, 7:06 PM

Even when I lower this slider

Ssmoke @Dr. Furkan Gözükara I am having a OOM error in Forge, on Massed Compute, do you...

Furkan Gözükara SECourses•8/27/24, 7:09 PM

can you edit this file and restart let me tell you

Furkan Gözükara SECourses•8/27/24, 7:09 PM

but before that actually can you dm me your entire screen screenshot

Furkan Gözükara SECourses•8/27/24, 7:09 PM

so i can see your exact config

Furkan Gözükara SECourses•8/27/24, 7:09 PM

because today i tested i have no error

smoke•8/27/24, 7:09 PM

Sure thing

Dreamer@TDFilmStudio•8/27/24, 7:33 PM

@Dr. Furkan Gözükara Is it possible to integrate joycaption as the default captioner for Supir, instead of llava?

nicearoni•8/27/24, 7:49 PM

running a big grid right now comparing different versions of a person lora... 25, 50, and 75 epoch versions all have long hair for "man as a cosmic jazz improviser whose music causes spontaneous formation of new elements and life forms and wearing eyeglasses" even though short hair in all training images

DDreamer@TDFilmStudio @Dr. Furkan Gözükara Is it possible to integrate joycaption as the default capti...

Furkan Gözükara SECourses•8/27/24, 7:49 PM

i may integrate

Nnicearoni running a big grid right now comparing different versions of a person lora... 25...

Furkan Gözükara SECourses•8/27/24, 7:49 PM

FFurkan Gözükara SECourses 😄

Dreamer@TDFilmStudio•8/27/24, 7:53 PM

Since you've mastered custom LoRA training for people, I encourage you to explore Flux LoRA training for products, objects, and art styles.

DDreamer@TDFilmStudio Since you've mastered custom LoRA training for people, I encourage you to explor...

Furkan Gözükara SECourses•8/27/24, 7:55 PM

yep it is also in progress

DDreamer@TDFilmStudio Since you've mastered custom LoRA training for people, I encourage you to explor...

fadounav•8/27/24, 7:55 PM

I agree, I don't understand why I'm losing the comic style I trained my lora with via comfyUI. I do have my avatar but in a realistic style, whereas I keep the comic side when I go through automatic 1111. I'd like to train a style to have consistency.

Ffadounav I agree, I don't understand why I'm losing the comic style I trained my lora wit...

Dreamer@TDFilmStudio•8/27/24, 7:57 PM

Flux is clearly excelling at art style and product LoRAs, as evidenced by the recent surge of these types on Civitai. Also I look forward to seeing architectural and other categories too.

Ffadounav I agree, I don't understand why I'm losing the comic style I trained my lora wit...

Furkan Gözükara SECourses•8/27/24, 7:59 PM

i added new config

Furkan Gözükara SECourses•8/27/24, 7:59 PM

Best_Configs_Better_Colors

Furkan Gözükara SECourses•8/27/24, 7:59 PM

it may be fix for style training

Furkan Gözükara SECourses•8/27/24, 7:59 PM

i will do more research with it not compoleted yet

Furkan Gözükara SECourses•8/27/24, 8:00 PM

@NicB we got error with forge

Furkan Gözükara SECourses•8/27/24, 8:00 PM