Software Engineering Courses (SECourses)•3y ago

some pictures look like your face is photoshopped inside the suit

felixsanzOP•6/6/23, 4:58 PM

but yeah, good quality

Le•6/6/23, 5:02 PM

I noticed you got full body results from just using your face to train the LoRA. Do I need full body pictures to have a LoRA that respects prompt?

Original message was deleted

ashley•6/6/23, 5:02 PM

Don't you have to adjust the width and height from 512x512 for this?

Ffelixsanz some pictures look like your face is photoshopped inside the suit

Furkan Gözükara SECourses•6/6/23, 5:02 PM

i will use AI to sort

Aashley Don't you have to adjust the width and height from 512x512 for this?

Le•6/6/23, 5:03 PM

it is 512x513 for a missclick. But resolution doesn't impact anything. As you can see, I have full body pictures using the same resolution. If I use any textual inversion, I also get good results.

Le•6/6/23, 5:03 PM

No results on my LoRA

Original message was deleted

Furkan Gözükara SECourses•6/6/23, 5:03 PM

for distant shots you need either higher resolution

Le•6/6/23, 5:03 PM

I got TONS of face pictures using (((full body:1.3))) on the prompt

Furkan Gözükara SECourses•6/6/23, 5:03 PM

or fixing with inpaint

Furkan Gözükara SECourses•6/6/23, 5:03 PM

did you have distant shots in your training data

FFurkan Gözükara SECourses or fixing with inpaint

Le•6/6/23, 5:03 PM

I tried that too. I get weird results.

Furkan Gözükara SECourses•6/6/23, 5:05 PM

tried inpaint?

Le•6/6/23, 5:05 PM

this was inpaint

Le•6/6/23, 5:05 PM

none was blonde, none looked like me

felixsanzOP•6/6/23, 5:05 PM

but inpaint using "only masked"

Furkan Gözükara SECourses•6/6/23, 5:05 PM

latest version of kohya didnt learned

Furkan Gözükara SECourses•6/6/23, 5:06 PM

maybe you used that version

FFurkan Gözükara SECourses latest version of kohya didnt learned

Le•6/6/23, 5:07 PM

what should I do then? there is no way to use Kohya with a downgraded version, right?

LLe what should I do then? there is no way to use Kohya with a downgraded version, r...

Furkan Gözükara SECourses•6/6/23, 5:08 PM

there is a way certainly

Furkan Gözükara SECourses•6/6/23, 5:08 PM

https://github.com/FurkanGozukara/Stable-Diffusion/blob/main/Tutorials/Generate-Studio-Quality-Realistic-Photos-By-Kohya-LoRA-Stable-Diffusion-Training-Full-Tutorial.md

GitHub

Stable-Diffusion/Generate-Studio-Quality-Realistic-Photos-By-Kohya-...

Best Stable Diffusion and AI Tutorials, Guides, News, Tips and Tricks - Stable-Diffusion/Generate-Studio-Quality-Realistic-Photos-By-Kohya-LoRA-Stable-Diffusion-Training-Full-Tutorial.md at main · ...

Furkan Gözükara SECourses•6/6/23, 5:08 PM

do a new git clone

Furkan Gözükara SECourses•6/6/23, 5:08 PM

then do

Furkan Gözükara SECourses•6/6/23, 5:08 PM

git checkout 63657088f4c35a376dd8a936f53e9b9a3b4b1168

Furkan Gözükara SECourses•6/6/23, 5:08 PM

this is the version i used

Furkan Gözükara SECourses•6/6/23, 5:08 PM

that is the version i used in the video

Le•6/6/23, 5:09 PM

Ok, I will try.

Le•6/6/23, 5:09 PM

Thank you

Furkan Gözükara SECourses•6/6/23, 5:09 PM

which gpu do you have?

Ffelixsanz but inpaint using "only masked"

Le•6/6/23, 5:09 PM

I tried that too. I get a funko pop on inpaint

Le•6/6/23, 5:09 PM

:crying:

Le•6/6/23, 5:11 PM

The quality is impressive. I got amazing results for my face. But the LoRA is unresponsive. It doesn't respect prompt. If it does I get a saturated low-quality version of myself and ignores full-body requests. It is so annoying.

Le•6/6/23, 5:11 PM

Going to try that tutorial.

LLe The quality is impressive. I got amazing results for my face. But the LoRA is un...

felixsanzOP•6/6/23, 5:12 PM

probably overtrained

Ffelixsanz probably overtrained

Le•6/6/23, 5:12 PM

I used something likee 2000 steps

felixsanzOP•6/6/23, 5:13 PM

holy s, maybe stable diffusion goes crazy with so much beauty xd

felixsanzOP•6/6/23, 5:13 PM

i don't like loras, they are not flexible at all in my experience