Software Engineering Courses (SECourses)•17mo ago

@Dr. Furkan Gözükara have you tried this prompt Anime illustration portrait of ohwx man, forest with

@Dr. Furkan Gözükara have you tried this prompt Anime illustration portrait of ohwx man, forest with lake in the background, high quality, beautiful face, action pose, clear focus, detailed and fantasy setting

dsienra•8/29/24, 3:07 PM

@Dr. Furkan Gözükara but I have to say same dataset same caption including class on sdxl no bleeding at all, I trained 20 or more people at the same time on a single sdxl lora and no bleeding at all, something is missing, idk

Ppixelscience Dude on reddit suposedly fine tuned it. Link is a discord link that doesn't work...

MaxiviOP•8/29/24, 3:08 PM

this is not a finetune, this is lora training, i went on his discord its just 512x512 lora training with fal

pixelscience•8/29/24, 3:08 PM

lol, I figured it might be something shady

MaxiviOP•8/29/24, 3:09 PM

his discord is quite shady too i left

Ddsienra yes I did, no class, now I balanced better the dataset, and I will make a test w...

Furkan Gözükara SECourses•8/29/24, 3:10 PM

balance matters

MMaxivi @Dr. Furkan Gözükara have you tried this prompt Anime illustration portrait of o...

Furkan Gözükara SECourses•8/29/24, 3:10 PM

it doesnt work

Furkan Gözükara SECourses•8/29/24, 3:11 PM

currently loras are overfitting

Ddsienra @Dr. Furkan Gözükara but I have to say same dataset same caption including class...

Furkan Gözükara SECourses•8/29/24, 3:11 PM

maybe due to flux nature

Furkan Gözükara SECourses•8/29/24, 3:11 PM

it needs more research

MMaxivi this is not a finetune, this is lora training, i went on his discord its just 51...

Furkan Gözükara SECourses•8/29/24, 3:11 PM

lol

Furkan Gözükara SECourses•8/29/24, 3:11 PM

512 is way lower quality people really doesnt test

Furkan Gözükara SECourses•8/29/24, 3:11 PM

you have to test with exact same configuration and compare

Furkan Gözükara SECourses•8/29/24, 3:12 PM

i will prepare full fine tuning guide hopefully with Kohya

Furkan Gözükara SECourses•8/29/24, 3:12 PM

also for flux you can have very different angles in train dataset

Furkan Gözükara SECourses•8/29/24, 3:12 PM

like looking down

Furkan Gözükara SECourses•8/29/24, 3:12 PM

looking up

Furkan Gözükara SECourses•8/29/24, 3:13 PM

completey side view

Furkan Gözükara SECourses•8/29/24, 3:13 PM

it really does amazing job

FFurkan Gözükara SECourses it needs more research

dsienra•8/29/24, 3:13 PM

may be, sdxl was mastered to its limits, some things about flux are still a mystery

Ddsienra may be, sdxl was mastered to its limits, some things about flux are still a myst...

Furkan Gözükara SECourses•8/29/24, 3:18 PM

yep so true

daroth•8/29/24, 3:19 PM

traning people/characters seem to be figured out mostly now, I'm curious, whats next? using the thousands of reg images to further finetune Flux dev?

Ddaroth traning people/characters seem to be figured out mostly now, I'm curious, whats ...

Furkan Gözükara SECourses•8/29/24, 3:20 PM

i will do fine tuning with my reg images

Furkan Gözükara SECourses•8/29/24, 3:20 PM

also a style training

Furkan Gözükara SECourses•8/29/24, 3:20 PM

will share style dataset too

Furkan Gözükara SECourses•8/29/24, 3:20 PM

and also then there is single image training challange :d

Furkan Gözükara SECourses•8/29/24, 3:20 PM

but also there is fine tuning config / workflow

Furkan Gözükara SECourses•8/29/24, 3:20 PM

i bet it will be way better than lora

FFurkan Gözükara SECourses i bet it will be way better than lora

�

𝔸𝖨ΞӾ•8/29/24, 3:24 PM

https://tenor.com/view/hell-yeah-brother-hulk-hogan-nodding-hogan-gif-12315629892838477060

Tenor

FFurkan Gözükara SECourses i will do fine tuning with my reg images

dsienra•8/29/24, 3:24 PM

did you try training with regularization but using no class? all your trainings looked amazing but all man end up looking like you

dsienra•8/29/24, 3:28 PM

may be doing that you maintain resemblance linking you to the token ohwx only and not to the token man

dsienra•8/29/24, 3:36 PM

I will watch your video now!!!!

novsky•8/29/24, 3:45 PM

is Q8 gguf flux model worse than Q5? becuse I have speed boost on Q8 and don't understand why

Ddsienra I will watch your video now!!!!

Furkan Gözükara SECourses•8/29/24, 3:58 PM

awesome

Nnovsky is Q8 gguf flux model worse than Q5? becuse I have speed boost on Q8 and don't u...

Furkan Gözükara SECourses•8/29/24, 3:59 PM

here comparison :

Furkan Gözükara SECourses•8/29/24, 3:59 PM

this is made for LLMs but same logic

Ddsienra did you try training with regularization but using no class? all your trainings ...

Furkan Gözükara SECourses•8/29/24, 4:00 PM

so how am i gonna this ? so it will be ohwx token of me, and man token for reg images right?

Furkan Gözükara SECourses•8/29/24, 4:00 PM

good idea i can test

Furkan Gözükara SECourses•8/29/24, 4:00 PM

but the issue is flux itself has inner captioning

Furkan Gözükara SECourses•8/29/24, 4:00 PM

so it still caption me as man :d

FFurkan Gözükara SECourses here comparison :

novsky•8/29/24, 4:02 PM

if it's only size issue, and speed boosts almost twice (6,3s to 3,7s/it) I will just need bigger ssd)))

novsky•8/29/24, 4:03 PM

thx for explaining

Dreamer@TDFilmStudio•8/29/24, 6:06 PM

@Dr. Furkan Gözükara After seeing your tutorial I am confused. We dont need to caption our training images for flux lora, right? (To achieve better quality)