Software Engineering Courses (SECourses)•13mo ago

Hey, in full stable diff video tutorial I've found that I have learning rate set to 8e-06 with your

Hey, in full stable diff video tutorial I've found that I have learning rate set to 8e-06 with your presets, whereas on video you have 1e-05. However I've used fast preset as I have 24 of VRAM. Is it the case?

AAleksAdv Hey, in full stable diff video tutorial I've found that I have learning rate set...

Furkan Gözükara SECourses•11/24/24, 10:24 PM

i updated them later

Furkan Gözükara SECourses•11/24/24, 10:24 PM

according to the newest libraries

Furkan Gözükara SECourses•11/24/24, 10:24 PM

on patreon

Furkan Gözükara SECourses•11/24/24, 10:25 PM

i am updating flux downloader for flux tools

AleksAdvOP•11/24/24, 10:37 PM

btw. what is resolution of your posture pictures you use for training? I've noticed that man dataset has 1024x1024, so are your photos also on 1024x1024? Also should be these as jpg or can be as png? Note that png is 10 times bigger.

AAleksAdv btw. what is resolution of your posture pictures you use for training? I've noti...

Furkan Gözükara SECourses•11/24/24, 10:43 PM

i used all 1024x1024

Furkan Gözükara SECourses•11/24/24, 10:43 PM

can be jpg or png both good

AleksAdvOP•11/24/24, 10:44 PM

ok, I've wrote python script to normalize to jpg and upscale/crop to 1024x1024 keeping the ratio

AleksAdvOP•11/24/24, 10:46 PM

another question, in your concept dataset you set prompt source as single file pointing to ohwx man.txt, but you have captions for each of the file, is the onetrainer looks at this file and also checks other txt files which coresponds to jpg files?

AleksAdvOP•11/24/24, 10:49 PM

I've also prefixed each of the caption txt file with recommended token, so for my this is "double" work

AleksAdvOP•11/24/24, 11:00 PM

also I'm testing now mask creation and after hitting "Create Masks" I see something being processed but after all I'm not getting any masks

AleksAdvOP•11/24/24, 11:00 PM

not in folder nor in preview

Sipriyani•11/24/24, 11:03 PM

@Furkan Gözükara SECourses by any chance have you tried, recent flux tools in swarm that is called? FLUX.1-Redux-dev??

AleksAdvOP•11/24/24, 11:15 PM

Mask works for Rembg, but is getting also body, not only head

SSipriyani @Furkan Gözükara SECourses by any chance have you tried, recent flux tools in s...

Furkan Gözükara SECourses•11/24/24, 11:23 PM

i am working on that right now

Furkan Gözükara SECourses•11/24/24, 11:23 PM

adding them to our unified downloader

AleksAdvOP•11/24/24, 11:42 PM

OK, regarding to Mask, it is a bug in OneTrainer, I've asked at their Discord channel

Furkan Gözükara SECourses•11/25/24, 12:49 AM

updating downloader

Fabricatedgirls•11/25/24, 1:15 AM

does kohya let you train flux loras within the interface yet or is it still scripts you gotta run manually

druhl•11/25/24, 3:53 AM

The GUI suffices?
May need the script for advanced stuff.

FFurkan Gözükara SECourses Click to see attachment

VioletAnt•11/25/24, 9:36 AM

Why does he not exactly look like trump? Is that difficult or even intended?

VVioletAnt Why does he not exactly look like trump? Is that difficult or even intended?

XCogni•11/25/24, 9:40 AM

Yep looks like a blend of the rock and trump

FFurkan Gözükara SECourses family stuff taking time 😄 today i worked entire day for the wife 😄

VioletAnt•11/25/24, 10:02 AM

hehe

I am wondering now do you have relatives also approaching you to generate images of them for their Instagram?

VVioletAnt Why does he not exactly look like trump? Is that difficult or even intended?

Furkan Gözükara SECourses•11/25/24, 10:38 AM

ye the experiments didnt work good

Furkan Gözükara SECourses•11/25/24, 10:38 AM

wasted so much time :d

VVioletAnt hehe 😄 I am wondering now do you have relatives also approaching you to generat...

Furkan Gözükara SECourses•11/25/24, 10:38 AM

yep few did :d

Furkan Gözükara SECourses•11/25/24, 10:38 AM

working on flux tools tutorial

FFurkan Gözükara SECourses ye the experiments didnt work good

XCogni•11/25/24, 10:41 AM

experiment dead?

XXCogni experiment dead?

Furkan Gözükara SECourses•11/25/24, 10:41 AM

it doesnt worth it

Furkan Gözükara SECourses•11/25/24, 10:41 AM

you can see all model trainings here : https://huggingface.co/MonsterMMORPG/Fix_FLUX_Bleeding_Experiments_with_DOP/tree/main

MonsterMMORPG/Fix_FLUX_Bleeding_Experiments_with_DOP at main

XCogni•11/25/24, 10:42 AM

Thats a lot of training

XCogni•11/25/24, 10:42 AM

have you tried with the dev distilled?

XXCogni have you tried with the dev distilled?

Furkan Gözükara SECourses•11/25/24, 10:43 AM

yes i tried

Furkan Gözükara SECourses•11/25/24, 10:43 AM

not the new method

Furkan Gözükara SECourses•11/25/24, 10:43 AM

but regular training

Furkan Gözükara SECourses•11/25/24, 10:43 AM

it also bleeds

XCogni•11/25/24, 10:44 AM

there is possibly a workaround though its a lot of effort

XCogni•11/25/24, 10:46 AM

you'd need to photobash yourself next to random men, and then caption it in details

XXCogni you'd need to photobash yourself next to random men, and then caption it in deta...

Furkan Gözükara SECourses•11/25/24, 11:00 AM

yes it helps significantly especially couple training

Furkan Gözükara SECourses•11/25/24, 11:59 AM

something so wrong with outpainting

Furkan Gözükara SECourses•11/25/24, 12:14 PM

doing more research

Furkan Gözükara SECourses•11/25/24, 12:38 PM

new unified downloader download speed

Furkan Gözükara SECourses•11/25/24, 12:56 PM

doing some revision test

Sebastianblood•11/25/24, 1:02 PM

Is it possible to customize the UI for forge ? Like adding brand logo, rearranging features

Not to delete any existing features, just to rearrange them so that the whole visual look is less complex and has personal visual taste @Furkan Gözükara SECourses

SSebastianblood Is it possible to customize the UI for forge ? Like adding brand logo, rearrangi...

Furkan Gözükara SECourses•11/25/24, 1:07 PM

i dont think

Furkan Gözükara SECourses•11/25/24, 1:25 PM

ViolA_ChaoS•11/25/24, 2:59 PM

Hey everyone! I've had mixed experiences with different Flux models and I'd like to share my findings and ask for help:

Successfully trained a LoRA using flux1-dev.safetensors with Dreambooth following the tutorial video. The results were excellent - I can generate both realistic and stylized images with great quality. However, the pose control with this model is quite poor.

When Flux tools were released, I was excited but encountered two issues:

When I try to use my previously trained safetensors (trained with flux1-dev) with the Flux tools, it doesn't work well - generates very poor quality images.
So I tried training a new Dreambooth LoRA using flux1-canny-dev.safetensors, but got this error:RuntimeError: Error(s) in loading state_dict for Flux:size mismatch for img_in.weight: copying a param with shape torch.Size([3072, 128]) from checkpoint, the shape in current model is torch.Size([3072, 64]).

Has anyone successfully trained a Dreambooth LoRA with flux1-canny-dev.safetensors? If so, could you share your setup/approach?

I really want to combine the quality I achieved with flux1-dev with proper pose control from the new Flux tools.

AAleksAdv btw. what is resolution of your posture pictures you use for training? I've noti...

ViolA_ChaoS•11/25/24, 3:13 PM

I have achieved much better results when using these resolutions. Often you can miss the point of the image if you crop it and use 1024x1024. But it is very important that if you use a size that is not 1024x1024 you crop at least 8 images to that size.

VViolA_ChaoS I have achieved much better results when using these resolutions. Often you can ...

Furkan Gözükara SECourses•11/25/24, 4:17 PM

makes sense