Software Engineering Courses (SECourses)•13mo ago

I've noticed earlier there is huge influence of aspect ratio on inference in FLUX dev, such as 3:2 h

I've noticed earlier there is huge influence of aspect ratio on inference in FLUX dev, such as 3:2 horizontal would give you professional photo incline, 16:9 - movie/series/ film grain, 2:1 vertical- phone / social network alike etc

TimsonOP•12/14/24, 2:53 PM

It's even more prominent than direct prompt

TTimson I've noticed earlier there is huge influence of aspect ratio on inference in FLU...

TimsonOP•12/14/24, 3:01 PM

this also got me thinking that comparing checkpoints with only single aspect ratio output somewhat limits it's scope, but not sure if making a grid with randomised AR is possible

TTimson Another noob question about training datasets: if I have mixed quality dataset, ...

Furkan Gözükara SECourses•12/14/24, 3:17 PM

i would say test all cases and see how they perform

Furkan Gözükara SECourses•12/14/24, 3:17 PM

hard to tell

FFurkan Gözükara SECourses i would say test all cases and see how they perform

TimsonOP•12/14/24, 3:21 PM

but there is no fundamental problem on training with mixed images sizes in a set?

Dazzastrous [4090]•12/14/24, 4:21 PM

Could have used your airship Doc from Cog

cyberbol•12/14/24, 4:35 PM

What is it new option ReActor ?

TimsonOP•12/14/24, 6:14 PM

Is it normal that saving flux dev fp16 checkpoint takes 6 minutes?
I've tested the drive on that remote to have above 500MB/s writing speed, I don't see any CPU or GPU load during this operation, what could be the bottleneck?

TimsonOP•12/14/24, 6:14 PM

what is the "normal" duration of this operation?

Latentnaut•12/14/24, 6:35 PM

Hello group, do you think a finetune training of this product will capture all these little letters and details?

TTimson Is it normal that saving flux dev fp16 checkpoint takes 6 minutes? I've tested t...

TimsonOP•12/14/24, 6:45 PM

figured this out, I cheeped out on 16GB RAM + 16 GB swap that gets 99% filled during checkpoint save. Lesson learned, seems you need 32GB RAM for seamless training :heher:

Nneo do we got any alternatives like Massed Compute? other than runpod

Renzo•12/14/24, 7:38 PM

I have used https://www.hyperstack.cloud/. They have reasonable prices. But you have to configure everything mostly from scratch.

TTimson but there is no fundamental problem on training with mixed images sizes in a set...

Furkan Gözükara SECourses•12/14/24, 11:31 PM

there shouldnt be

Ccyberbol What is it new option ReActor ?

Furkan Gözükara SECourses•12/14/24, 11:31 PM

face swap

TTimson Is it normal that saving flux dev fp16 checkpoint takes 6 minutes? I've tested t...

Furkan Gözükara SECourses•12/14/24, 11:32 PM

no savings way faster for me

Furkan Gözükara SECourses•12/14/24, 11:32 PM

like few seconds

LLatentnaut Hello group, do you think a finetune training of this product will capture all t...

Furkan Gözükara SECourses•12/14/24, 11:32 PM

not likely

Furkan Gözükara SECourses•12/14/24, 11:32 PM

some text really small

TTimson figured this out, I cheeped out on 16GB RAM + 16 GB swap that gets 99% filled du...

Furkan Gözükara SECourses•12/14/24, 11:32 PM

actually get 64 gb

Furkan Gözükara SECourses•12/15/24, 12:09 AM

Furkan Gözükara SECourses•12/15/24, 1:12 AM

published : https://www.patreon.com/posts/117990364

Patreon

MMAudio 1-Click Installers for Windows, RunPod and Massed Compute -...

Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon

Furkan Gözükara SECourses•12/15/24, 1:12 AM

Sipriyani•12/15/24, 1:18 AM

and interesting method for , consistent images and reposing ... however, each scene will need a lora training. https://www.youtube.com/watch?v=MbQv8zoNEfY

Furkan Gözükara SECourses•12/15/24, 1:32 AM

best consistent way is training

Furkan Gözükara SECourses•12/15/24, 1:32 AM

you can do multi step to get what you want

QM•12/15/24, 1:33 AM

I have a 6x 4090 setup on Win 11 - is there a way to use all GPUs for finetune training ?

QQM I have a 6x 4090 setup on Win 11 - is there a way to use all GPUs for finetune t...

Furkan Gözükara SECourses•12/15/24, 1:34 AM

nope

Furkan Gözükara SECourses•12/15/24, 1:34 AM

fine tuning need 80 gb gpus

Furkan Gözükara SECourses•12/15/24, 1:34 AM

but you can do 6 parallel training

Furkan Gözükara SECourses•12/15/24, 1:35 AM

i got to sleep now hopefully see you tomorrow

QM•12/15/24, 1:35 AM

is it possible to build up a finetuned or LoRA model - i.e. pictures of me, wife kids etc?

QQM is it possible to build up a finetuned or LoRA model - i.e. pictures of me, wife...

Furkan Gözükara SECourses•12/15/24, 1:38 AM

yes

Furkan Gözükara SECourses•12/15/24, 1:38 AM

have all of you in same pictures as well during training

Furkan Gözükara SECourses•12/15/24, 1:38 AM

and caption them

QM•12/15/24, 1:38 AM

have a good night

Furkan Gözükara SECourses•12/15/24, 1:40 AM

like ohwx man bbuk woman qwxx kid

Furkan Gözükara SECourses•12/15/24, 1:40 AM

i didnt test this but logic is like this

Furkan Gözükara SECourses•12/15/24, 1:40 AM

FFurkan Gözükara SECourses like ohwx man bbuk woman qwxx kid

QM•12/15/24, 1:40 AM

thank you

FFurkan Gözükara SECourses but you can do 6 parallel training 😄

QM•12/15/24, 1:54 AM

I am curious to maximise the PCs ability - with 6x 4090 and 512 DDR5 RAM (PCIe 5) - to train models faster

FFurkan Gözükara SECourses you can do multi step to get what you want 🙂

Sipriyani•12/15/24, 1:57 AM

doctor what exactly you mean with multi step?

Sipriyani•12/15/24, 1:59 AM

also another question, can I re-finetune, a flux model file which is already fine tuned?

dsienra•12/15/24, 3:54 AM

https://civitai.com/models/1034630/hunyuan-video-arnold-schwarzenegger-lora?modelVersionId=1160430 Great hunyuan lora trained only with 15 images, this video model generalize very well. very promising

QQM I am curious to maximise the PCs ability - with 6x 4090 and 512 DDR5 RAM (PCIe 5...

XCogni•12/15/24, 10:16 AM

sell the 6 4090s, buy 2 6000adas, or wait for the new blackwell chips coming out

FFurkan Gözükara SECourses like few seconds

TimsonOP•12/15/24, 11:24 AM

and what is you time for swapping fine-tuned full checkpoints in SwarmUI? Mine like 120s per switch

There is this lines in debug log during model switch are they ok? clip_l is loaded and working during generation, and it says something about casting although I'm saving checkpoints as fp16