Software Engineering Courses (SECourses)•17mo ago

Raw vs eyeglasses prompt - it is interesting that FLUX is learning me without eyeglasses when not pr

Raw vs eyeglasses prompt - it is interesting that FLUX is learning me without eyeglasses when not prompted. This was never case with SDXL

FFurkan Gözükara SECourses Raw vs eyeglasses prompt - it is interesting that FLUX is learning me without ey...

�

𝔸𝖨ΞӾ•8/21/24, 12:21 AM

Perhaps due to captioning with the dataset?

�𝔸𝖨ΞӾ Perhaps due to captioning with the dataset?

Furkan Gözükara SECoursesOP•8/21/24, 12:22 AM

same i just use ohwx man

Furkan Gözükara SECoursesOP•8/21/24, 12:22 AM

test reg images impact too

Furkan Gözükara SECoursesOP•8/21/24, 12:24 AM

running a big test for an hard prompt

�

𝔸𝖨ΞӾ•8/21/24, 12:32 AM

oh ok yeah that makes sense now. I just figured probably one of your training images didn't have the glasses described in the caption, but then again the frames of the glasses aren't super bold and noticeable from a distance so that could be another reason

Darth Teethius•8/21/24, 12:36 AM

does anyone know if normal alternating via prompt such as [x|y] where it alternates between x and y each step is still usable in a natural language style model like flux?

sergio_gonzo•8/21/24, 12:49 AM

maybe i will wait more, so the process can make it with a gpu 16 vram, maybe onetrainer make support for flux

�𝔸𝖨ΞӾ oh ok yeah that makes sense now. I just figured probably one of your training im...

Furkan Gözükara SECoursesOP•8/21/24, 1:09 AM

all images has eyeglasses and only ohwx man :d

DDarth Teethius does anyone know if normal alternating via prompt such as [x|y] where it alterna...

Furkan Gözükara SECoursesOP•8/21/24, 1:10 AM

flux has wildcards

Ssergio_gonzo maybe i will wait more, so the process can make it with a gpu 16 vram, maybe one...

Furkan Gözükara SECoursesOP•8/21/24, 1:10 AM

yep

Furkan Gözükara SECoursesOP•8/21/24, 1:10 AM

if kohya fixes also

Furkan Gözükara SECoursesOP•8/21/24, 1:18 AM

eular normal vs karras

Furkan Gözükara SECoursesOP•8/21/24, 1:18 AM

testing all combinations

Furkan Gözükara SECoursesOP•8/21/24, 1:25 AM

karras make huge diff and make it more real

Furkan Gözükara SECoursesOP•8/21/24, 1:25 AM

will post entire grid on patreon

Sipriyani•8/21/24, 1:25 AM

cant wait the config files and tutorials.

SSipriyani cant wait the config files and tutorials. 🤓

Furkan Gözükara SECoursesOP•8/21/24, 1:26 AM

adding configs already :d

Sipriyani•8/21/24, 1:29 AM

now just seen, from all the png kpeg attachments that there is one zip file, .. I downloaded.
I have the latest branch kohya.
now i will start the training.

Sipriyani•8/21/24, 1:30 AM

any recommendation for 4090 and beginner?

Sipriyani•8/21/24, 1:32 AM

1e_04_bf16_128_rank_full_bf16 ? maybe?

SSipriyani 1e_04_bf16_128_rank_full_bf16 ? maybe?

Furkan Gözükara SECoursesOP•8/21/24, 1:41 AM

Look very top written best ones

FFurkan Gözükara SECourses karras make huge diff and make it more real

Haashim•8/21/24, 1:50 AM

What guidance you using

FFurkan Gözükara SECourses eular normal vs karras

Haashim•8/21/24, 1:51 AM

No more Motorcycle handles on the T rex back now lol

HHaashim What guidance you using

Furkan Gözükara SECoursesOP•8/21/24, 1:52 AM

3.5 still gonna test it too

HHaashim No more Motorcycle handles on the T rex back now lol

Furkan Gözükara SECoursesOP•8/21/24, 1:53 AM

dsienra•8/21/24, 2:03 AM

In my first tests I'm still having problems whith concept bleeding

dsienra•8/21/24, 2:04 AM

works fine if you train just one person

Furkan Gözükara SECoursesOP•8/21/24, 2:05 AM

currently impact of lora so powrful

Furkan Gözükara SECoursesOP•8/21/24, 2:05 AM

it mixed me even with woman prompt

dsienra•8/21/24, 2:05 AM

if you train many people mix them all, is not paying attention to the captions

dsienra•8/21/24, 2:06 AM

mix person 1 with person two the problem is worse if they are from the same class

dsienra•8/21/24, 2:07 AM

must to be the text ecoder

dsienra•8/21/24, 2:08 AM

the same issue I hade with sd3

FFurkan Gözükara SECourses currently impact of lora so powrful

dsienra•8/21/24, 2:09 AM

yes is only training the unet the text captions are ignored

RMK•8/21/24, 2:09 AM

@Dr. Furkan Gözükara

Are you doing FLUX Lora training or dream booth training?

And do you need Adetailer for generating images? Is this inside the SwarmUI?

Ddsienra yes is only training the unet the text captions are ignored

Furkan Gözükara SECoursesOP•8/21/24, 2:14 AM

yep

RRMK @Dr. Furkan Gözükara Are you doing FLUX Lora training or dream booth training?...

Furkan Gözükara SECoursesOP•8/21/24, 2:14 AM

flux lora

dsienra•8/21/24, 2:14 AM

or looks like the text captions are ignored

Furkan Gözükara SECoursesOP•8/21/24, 2:14 AM

i use segment feature of swarmui

Furkan Gözükara SECoursesOP•8/21/24, 2:15 AM

https://youtu.be/HKX8_F1Er_w

YouTubeSECourses

Zero to Hero Stable Diffusion 3 Tutorial with Amazing SwarmUI SD We...

Do not skip any part of this tutorial to master how to use Stable Diffusion 3 (SD3) with the most advanced generative AI open source APP SwarmUI. Automatic1111 SD Web UI or Fooocus are not supporting the #SD3 yet. Therefore, I am starting to make tutorials for SwarmUI as well. #StableSwarmUI is officially developed by the StabilityAI and your m...

Furkan Gözükara SECoursesOP•8/21/24, 2:29 AM

huge speed up for rtx 4xxx series for flux

Furkan Gözükara SECoursesOP•8/21/24, 2:29 AM

https://www.reddit.com/r/StableDiffusion/comments/1ex3ol5/comfyui_experimental_rtx_40_series_update/

From the StableDiffusion community on Reddit

Explore this post and more from the StableDiffusion community

FFurkan Gözükara SECourses yep

dsienra•8/21/24, 2:52 AM

one test that you can do is training without the class just "ohwx" leve the token man untouched, then trigger ohwx and see if is doing something or is ignored completely, if that its the case thats the reason why concepts bleed and is not learning anything from the captions

FFurkan Gözükara SECourses huge speed up for rtx 4xxx series for flux

Haashim•8/21/24, 3:15 AM

Isnt it late at home? haha

FFurkan Gözükara SECourses https://www.reddit.com/r/StableDiffusion/comments/1ex3ol5/comfyui_experimental_r...

Haashim•8/21/24, 3:15 AM

wonder if this is true on runpod 4090

virginia•8/21/24, 5:42 AM

@Dr. Furkan Gözükara possible to show me the json for sdxl lora settings

Ddsienra one test that you can do is training without the class just "ohwx" leve the toke...

clarkcheekykent•8/21/24, 6:34 AM

I trained 2 person lora (nmwx for myself and ohwx for my wife), all images of me tagged with nmwx, despite prompting simply "nmwx" it returned images of my wife. BUT its very consistent that all "man" return me and all "woman" prompts returns my wife. So it looks like the model just overfitted for all men to look like me and all women look like my wife.

Also "nmwx man, HUGO BOSS advertisement" and "man, HUGO BOSS advertisement" return 99.5% the same image and likeness, just minor details changed.

BTW the 2 character lora works amazing. no bleeding between the two characters for me and can generate both within one image. Probably because its two genders, i'm doubting it would work if its two men though without training the encoder