Software Engineering Courses (SECourses)•3y ago

you dont need to do git pull

Sspectral_flux I realized now it works, but now it launchs using 127.0.01:7860 instead of 0.0.0...

Furkan Gözükara SECoursesOP•5/1/23, 10:02 PM

probably you didnt kill previous instance

Mmayen Click to see attachment

Furkan Gözükara SECoursesOP•5/1/23, 10:02 PM

no first option

ZZono50 Click to see attachment

Furkan Gözükara SECoursesOP•5/1/23, 10:03 PM

face is so high qualty

FFurkan Gözükara SECourses you dont need to do git pull

spectral_flux•5/1/23, 10:03 PM

Oh I see, will try that! Thanks again

Zono50•5/1/23, 10:12 PM

Thanks doc

Zono50•5/1/23, 10:24 PM

that's using xformers 2

snek•5/1/23, 10:29 PM

@Furkan Gözükara SECourses have u tried using more augments apart from horizontal flip to see if they improve training?

Ssnek @Furkan Gözükara SECourses have u tried using more augments apart from horizonta...

Furkan Gözükara SECoursesOP•5/1/23, 10:29 PM

which ones for example?

snek•5/1/23, 10:30 PM

i was thinking this, idk if makes sense using it tho, but there are plenty more https://pytorch.org/vision/stable/generated/torchvision.transforms.RandomErasing.html#torchvision.transforms.RandomErasing

Ssnek i was thinking this, idk if makes sense using it tho, but there are plenty more ...

Furkan Gözükara SECoursesOP•5/1/23, 10:33 PM

i think trying to improve training dataset would have more dramatic effect

snek•5/1/23, 10:33 PM

like randomly grayscale, if its 5% of the times will not affect too much

snek•5/1/23, 10:33 PM

yeah i was thinking more in terms of preventing overfitting and still improving if possible

snek•5/1/23, 10:35 PM

iam trying random image rotation -45,45 , just for the sake of experimenting

Ssnek iam trying random image rotation -45,45 , just for the sake of experimenting

Furkan Gözükara SECoursesOP•5/1/23, 10:36 PM

thanks let us know the results please

snek•5/1/23, 10:36 PM

sure

Zono50•5/1/23, 10:38 PM

i did 256 using locon on kohya, seemed to get better results

crazygonuts•5/1/23, 10:39 PM

@Furkan Gözükara SECourses I'm interested to try and run deep floyd IF locally, I have a 3090 so i should have enough RAM. I thought you had mentioned the possibility of making a video for that?

Ccrazygonuts @Furkan Gözükara SECourses I'm interested to try and run deep floyd IF locally, ...

Furkan Gözükara SECoursesOP•5/1/23, 10:46 PM

yep it is on my list

Furkan Gözükara SECoursesOP•5/1/23, 10:46 PM

with a good ui

crazygonuts•5/1/23, 10:53 PM

I've read that voicevox is a good opensource AI text-to-voice model but also as far as i can tell its only for japanese at the moment. Is there any equivalent for english that anyone knows about?

FFurkan Gözükara SECourses no first option

mayen•5/1/23, 10:55 PM

after torch 2.o install

mayen•5/1/23, 10:55 PM

Mmayen Click to see attachment

Furkan Gözükara SECoursesOP•5/1/23, 11:11 PM

well maybe kohya ss is not as optimized as sd web ui

Furkan Gözükara SECoursesOP•5/1/23, 11:11 PM

because i know 8 gb vram people using sd web ui lora

rafraf•5/1/23, 11:16 PM

what is the correct way to make an image with two LoRA's: 1) the subject LoRA, training for face accuracy, and 2) an outfit LoRA, training for body and outfit accuracy. In both cases, I am training for unique real object. I suppose I should treat both like you did your suit portraits using LoRA with class images, no prompt. Am I correct to suppose that I should treat both face and outfit the same way? with no captions, using class images? @Furkan Gözükara SECourses @Zono50