Software Engineering Courses (SECourses)•14mo ago

hi <@205854764540362752> what is the best way to train multiple person or is it best to create a m

hi @Furkan Gözükara SECourses what is the best way to train multiple person or is it best to create a model for each person ? If second
option is better what is the best aproach in confy ui to generate a single image with multiple model?

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/29/24, 9:28 AM

If I have 15 great 2048x2048 face pics, should I just these for training, or must I downsize to 1024 first? Any chance the 2048 pics will give better result or not?

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/29/24, 11:01 AM

Another thing, if I want the best possible lora, I won't loose details by first making fine tuned checkpoints, then converting? For this character I really want both, so was thinking of just doing fine tune, then convert the best epochs to loras.

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/29/24, 11:12 AM

I now get error when using your kohya_v16 install script, can't find the gui.sh.

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/29/24, 11:13 AM

massedcompute

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/29/24, 11:14 AM

I figured out my error, I deleted the kohya ss folder, didn't think I needed that when I had your script

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•11/29/24, 11:17 AM

Foud kohya_ss on git and did a git clone inside apps folder, hope that works

Wwardensc2 I try to train new character checkpoint using pixelwave and failed miserably, th...

Furkan Gözükara SECourses•11/29/24, 12:24 PM

i used it

Furkan Gözükara SECourses•11/29/24, 12:24 PM

but pixelwave is overfit model

Ggio3000.hi <@205854764540362752> what is the best way to train multiple person or is i...

Furkan Gözükara SECourses•11/29/24, 12:24 PM

yes a model for each person

Furkan Gözükara SECourses•11/29/24, 12:24 PM

second question you can try regional prompting in swarmui

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭If I have 15 great 2048x2048 face pics, should I just these for training, or mus...

Furkan Gözükara SECourses•11/29/24, 12:24 PM

for face i think 1024 works best

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭I figured out my error, I deleted the kohya ss folder, didn't think I needed tha...

Furkan Gözükara SECourses•11/29/24, 12:25 PM

ye script uses base installation

Konstantinas•11/29/24, 1:47 PM

if training on de-destilled flux model in kohya, I guess prompts will simply fail to work?

KKonstantinas if training on de-destilled flux model in kohya, I guess prompts will simply fai...

Furkan Gözükara SECourses•11/29/24, 1:48 PM

what do you mean?

Furkan Gözükara SECourses•11/29/24, 1:48 PM

i trained on de-distilled

Furkan Gözükara SECourses•11/29/24, 1:48 PM

https://medium.com/@furkangozukara/thoroughly-experimented-with-fine-tuning-dreambooth-training-of-flux-dev-de-distill-pixelwave-13fab8c625bf

Medium

Thoroughly experimented with Fine-Tuning / DreamBooth training of F...

Thoroughly experimented with Fine-Tuning / DreamBooth training of Flux-dev-de-distill, PixelWave v03, Verus Vision and base FLUX Dev model…

Konstantinas•11/29/24, 1:49 PM

I mean, if I use "sample prompts" in kohya ss, it will think this is Flux model, but as I have de-distilled, this needs different CGF strategy? Its like, in Comfy UI, I need different workflow, if I use de-distilled model?

Konstantinas•11/29/24, 1:50 PM

you probably simply did not use sample prompting when training?

KKonstantinas I mean, if I use "sample prompts" in kohya ss, it will think this is Flux model,...

Furkan Gözükara SECourses•11/29/24, 2:04 PM

what changes is

Furkan Gözükara SECourses•11/29/24, 2:04 PM

instead of cfg 1

Furkan Gözükara SECourses•11/29/24, 2:04 PM

we use cfg 3.5

Furkan Gözükara SECourses•11/29/24, 2:04 PM

both training and inference

Konstantinas•11/29/24, 2:05 PM

I see, thanks. I'll disable "sample prompting" for now, seems like not yet way to supply it to kohya ss

Konstantinas•11/29/24, 2:06 PM

as only has "cfg scale" as an option there

KKonstantinas I see, thanks. I'll disable "sample prompting" for now, seems like not yet way t...

Furkan Gözükara SECourses•11/29/24, 2:14 PM

ye i dont recommend samples

Konstantinas•11/29/24, 2:15 PM

I just like to sneak-peak early on how it goes

but I guess I can just stop training at cehckpoints, render previews manually, and resume from state if needed

Zhigarev•11/29/24, 2:54 PM

Hello everyone! Owners of the RTX 3090 and CPUs (WITHOUT integrated graphics), has anyone managed to achieve the training speed 7.3 s/it stated in the config by @Furkan Gözükara SECourses ? Even when I don't turn anything off, I still end up using 1000GB of VRAM on Windows 11, even if absolutely no third-party programs are running. How it possible ? i got onaly 11.16 s/it with 24GB config file...

Konstantinas•11/29/24, 3:07 PM

about VRAM, I notice one thing. I have RTX 4090, on 24GB config, I have 6s/it; Then I swap blocks to 12GB config, and have 6.7s/it, almost no difference, but now I have lots of free VRAM and can use my PC normally alongside training. Have no idea why is that so

Konstantinas•11/29/24, 3:08 PM

maybe its because I have 128GB of CPU ram