Software Engineering Courses (SECourses)•15mo ago

@Dr. Furkan Gözükara New test Flux-De-Distilled with this settings the lora at 160 epochs looks seve

@Dr. Furkan Gözükara New test Flux-De-Distilled with this settings the lora at 160 epochs looks severely undertrained, may be Flux-De-Distilled requieres higher LR

dsienraOP•10/15/24, 4:06 PM

I test it on the de-distilled model and on regular Flux-Dev

dsienraOP•10/15/24, 4:08 PM

I'm going to resume the training with LR 0.0001

dsienraOP•10/15/24, 4:08 PM

and see if it converges

Ddsienra @Dr. Furkan Gözükara New test Flux-De-Distilled with this settings the lora at 1...

Furkan Gözükara SECourses•10/15/24, 4:15 PM

totally could be

Furkan Gözükara SECourses•10/15/24, 4:15 PM

i will research it later hopefully

Furkan Gözükara SECourses•10/15/24, 4:15 PM

i repoırted erroır to kohya : https://github.com/bmaltais/kohya_ss/issues/2901

GitHub

SD3 branch ERROR: Cannot install -r requirements.txt (line 1), -r r...

SD3 flux branch is broken when doing fresh install @kohya-ss @bmaltais Collecting wcwidth>=0.2.5 Using cached wcwidth-0.2.13-py2.py3-none-any.whl (34 kB) INFO: pip is looking at multiple version...

FFurkan Gözükara SECourses totally could be

dsienraOP•10/15/24, 4:18 PM

I gonna resume from 160 epochs at 0.001 for 10 epochs and test if get signs of converging, I will report

Furkan Gözükara SECourses•10/15/24, 4:53 PM

guys dont use 24gb massed compute instances - or get at least 2x GPU for bigger RAM

Furkan Gözükara SECourses•10/15/24, 4:53 PM

Vigilence•10/15/24, 5:43 PM

How to know what the vram is when selecting a gpu? Its not shown for most gpus.

VVigilence How to know what the vram is when selecting a gpu? Its not shown for most gpus.

Furkan Gözükara SECourses•10/15/24, 5:57 PM

best way is searching google

Ddsienra I gonna resume from 160 epochs at 0.001 for 10 epochs and test if get signs of c...

dsienraOP•10/15/24, 6:00 PM

@Dr. Furkan Gözükara now with lr at 0.0001 and 10 more epochs can see some resemblance with lora weight at 1.5, so is learning

Elwii•10/15/24, 7:19 PM

General question on dreambooth finetuning: When switching from 15 img to 150 img, do you still train as many steps as with 15 img or is it 10x more?

EElwii General question on dreambooth finetuning: When switching from 15 img to 150 img...

Furkan Gözükara SECourses•10/15/24, 7:37 PM

sadly no formula

Furkan Gözükara SECourses•10/15/24, 7:37 PM

for example 160*15 steps is best for 15 images

Furkan Gözükara SECourses•10/15/24, 7:37 PM

80 * 256 best for 256 images

Furkan Gözükara SECourses•10/15/24, 7:37 PM

batch size 1

FFurkan Gözükara SECourses 80 * 256 best for 256 images

Elwii•10/15/24, 7:38 PM

Thanks a lot

fadounav•10/15/24, 8:12 PM

Hey, is it now possible to train a style?

Ffadounav Hey, is it now possible to train a style?

Furkan Gözükara SECourses•10/15/24, 8:14 PM

it was always

Furkan Gözükara SECourses•10/15/24, 8:15 PM

here full info

Furkan Gözükara SECourses•10/15/24, 8:15 PM

https://huggingface.co/MonsterMMORPG/3D-Cartoon-Style-FLUX

MonsterMMORPG/3D-Cartoon-Style-FLUX · Hugging Face

Furkan Gözükara SECourses•10/15/24, 8:25 PM

you guys tested these 2 against fp16?

Furkan Gözükara SECourses•10/15/24, 8:25 PM

e4m3fn vs e5m2

Furkan Gözükara SECourses•10/15/24, 8:27 PM

i guess i will grid test now

Furkan Gözükara SECourses•10/15/24, 8:43 PM

ye i think e4m3fn better
https://imgsli.com/MzA3OTY0/0/2

Furkan Gözükara SECourses•10/15/24, 8:44 PM

32 GB RAM sucks so bad

FFurkan Gözükara SECourses it was always

fadounav•10/15/24, 8:53 PM

ok because you told me that (For stylized formations the flow is still not great) thank you very much

FFurkan Gözükara SECourses ye i think e4m3fn better https://imgsli.com/MzA3OTY0/0/2

fadounav•10/15/24, 8:54 PM

That's a good stuff to compare

Ffadounav ok because you told me that (For stylized formations the flow is still not great...

Furkan Gözükara SECourses•10/15/24, 8:54 PM

well if you want fully stylized person generation

Furkan Gözükara SECourses•10/15/24, 8:54 PM

you should train on stylized images

Furkan Gözükara SECourses•10/15/24, 8:55 PM

first do fine tuning

Furkan Gözükara SECourses•10/15/24, 8:55 PM

generate stylized images

Furkan Gözükara SECourses•10/15/24, 8:55 PM

retrain

Furkan Gözükara SECourses•10/15/24, 8:55 PM

by the way fine tuning yields way better stylization

Furkan Gözükara SECourses•10/15/24, 8:55 PM

than lora

fadounav•10/15/24, 8:57 PM

@Dr. Furkan Gözükara Yes, I'm at that stage, I've trained several stylized characters, but I would have liked to train the style in order to create object environments and the like. The aim is to create a world in the same style on Comfy UI.

fadounav•10/15/24, 8:59 PM

Like this

Ffadounav Like this

Furkan Gözükara SECourses•10/15/24, 9:05 PM

so if you can collect such images as dataset

Furkan Gözükara SECourses•10/15/24, 9:05 PM

it will work perfect

fadounav•10/15/24, 9:12 PM

@Dr. Furkan Gözükara Ok here's my theory, I train a Dreambooth with a character I call X for example, then I train this model with another character I call Y, always in the same style. With the captions well detailed, the goal is to have a final model that would allow me to make X and or Y appear via the prompt, is this consistent? or is it better to go through lora? btw I can have 256 images per character, if I win against my laziness lol

Ffadounav @Dr. Furkan Gözükara Ok here's my theory, I train a Dreambooth with a character ...

Furkan Gözükara SECourses•10/15/24, 9:16 PM

for SDXL it can work

Furkan Gözükara SECourses•10/15/24, 9:16 PM

for FLUX they bleed

Furkan Gözükara SECourses•10/15/24, 9:16 PM

also you may train all at once

Furkan Gözükara SECourses•10/15/24, 9:16 PM

assign differnt unique toknes

Furkan Gözükara SECourses•10/15/24, 9:16 PM

ohwx man

Furkan Gözükara SECourses•10/15/24, 9:16 PM