Software Engineering Courses (SECourses)•3y ago

possible to train for style with a couple hundred images? i saw the recent midjourney video uses tho

possible to train for style with a couple hundred images? i saw the recent midjourney video uses thousands of images for the style training... i tested with a few hundred and got decent results but i was not able to drive the output via the prompt (results looked good... but mostly just random images in my style regardless of prompt)

what should i change in my training to get better prompt weight in my outputs?

BBryan possible to train for style with a couple hundred images? i saw the recent midjo...

Furkan Gözükara SECourses•3/23/23, 8:06 PM

i think possible

Furkan Gözükara SECourses•3/23/23, 8:06 PM

today i will try with 274 images

Furkan Gözükara SECourses•3/23/23, 8:06 PM

to teach style

Furkan Gözükara SECourses•3/23/23, 8:06 PM

and 137 images to train myself

BryanOP•3/23/23, 8:06 PM

cool, will there be a video with the results?

BBryan possible to train for style with a couple hundred images? i saw the recent midjo...

Furkan Gözükara SECourses•3/23/23, 8:06 PM

to get more prompt weight probably more training

BBryan cool, will there be a video with the results?

Furkan Gözükara SECourses•3/23/23, 8:06 PM

Furkan Gözükara SECourses•3/23/23, 8:07 PM

i am working on corridor crew style video

Furkan Gözükara SECourses•3/23/23, 8:07 PM

animation

BryanOP•3/23/23, 8:07 PM

yeah i didn't train much since it takes so damn long. it was looking really good so i stopped the training, i guess i need to try much more

BBryan yeah i didn't train much since it takes so damn long. it was looking really good...

Furkan Gözükara SECourses•3/23/23, 8:09 PM

yep training takes huge time

Furkan Gözükara SECourses•3/23/23, 8:09 PM

when more images to train

Furkan Gözükara SECourses•3/23/23, 8:10 PM

i will post more classification images to patreon. aesthetic and photo of man

I A•3/23/23, 8:35 PM

Hello again, after a fresh install of A1111 and no xformers, I can confirm that everything is running fine, and loading faster than before. But I am not using A1111 for training. Dreambooth has to be installed manually, but it's OK.

I A•3/23/23, 8:36 PM

As to the training, I was able to get 60% of images that look like the subject trained, nothing as good as in your videos.

I A•3/23/23, 8:36 PM

When I do it with SD 2.1 or Protogen, it gets worse, no resemblance at all. Also the biggest problem in general is that if use toke + class I get some results, if I just do the usual "closeu portrait photo of...." I get other people!!!

GP•3/23/23, 8:38 PM

Perhaps a slight side step from current conversation, but I'm interested in training SD on some pixel art characters for use in a platform game. I've tried training it to understand that I want a white background around each image but I haven't had any luck. This was with textual inversion. Does anyone have any suggestions? (The image below is the kind of output I'm after, white background surrounding the subject)

I A•3/23/23, 8:40 PM

By the way Dr. Furkan, I still seem to have some confusion with numbers of steps/epochs, because the UI says "steps-epochs" and it's confusing. So I trained with values from 150-300 (whatever that is) and got models that go up to 12.000 (whatever that is). Could you please explain this in simple way?

I A•3/23/23, 8:41 PM

LR I'm using 0.000001 and not sure if go lower will be better. Like 1.0e-7 will that help?

II A By the way Dr. Furkan, I still seem to have some confusion with numbers of steps...

Furkan Gözükara SECourses•3/23/23, 9:03 PM

i explain these in videos actually

Furkan Gözükara SECourses•3/23/23, 9:03 PM

i will also mention in next video

I A•3/23/23, 9:13 PM

Yes, I saw all the videos, but it's OK. I did some research on the Internet, to clarify this further.

I A•3/23/23, 9:15 PM

What about the other problem? Using your exact settings, I can't work with prompts. token + class generates some images, then as soon as I write a prompt, I loose the subject. In your video that didn't happen. I'm using SD 1.5 and Protogen, same exact settings.

Stevens•3/23/23, 9:39 PM

I got the bot working on my server with my home install! I used a thing a dude put on github. Will not post the link unless/until doc approves but it's pretty cool.

II A What about the other problem? Using your exact settings, I can't work with promp...

Furkan Gözükara SECourses•3/23/23, 9:51 PM

something must be off. if you are our patreon supporter i can setup a training on your computer if you wish

SStevens I got the bot working on my server with my home install! I used a thing a dude ...

Furkan Gözükara SECourses•3/23/23, 9:51 PM

if anyone interested they can message you i think

Furkan Gözükara SECourses•3/23/23, 9:52 PM

i am starting a new training with newest dreambooth

Furkan Gözükara SECourses•3/23/23, 9:52 PM

in a minute

Stevens•3/23/23, 9:54 PM

Thanks, doc! What level patron gets troubleshooting help? I

Furkan Gözükara SECourses•3/23/23, 9:59 PM

/workspace/photo of man 6080
/workspace/aesthetic 10598
/workspace/training images 137
/workspace/style_512 274

SStevens Thanks, doc! What level patron gets troubleshooting help? I

Furkan Gözükara SECourses•3/23/23, 9:59 PM

well if it is easy basic level is sufficient . if you need longer help please get gold :d

Furkan Gözükara SECourses•3/23/23, 9:59 PM

time to train

Furkan Gözükara SECourses•3/23/23, 9:59 PM

Furkan Gözükara SECourses•3/23/23, 9:59 PM

multi concept at once

Szutykos Macska•3/23/23, 10:06 PM

Hello. I don't know who works on the youtube tutorials that led me here but he's a beautiful human beign.

SSzutykos Macska Hello. I don't know who works on the youtube tutorials that led me here but he's...

Furkan Gözükara SECourses•3/23/23, 10:06 PM

thank you . i am

Furkan Gözükara SECourses•3/23/23, 10:06 PM

welcome

Furkan Gözükara SECourses•3/23/23, 10:10 PM

caching latents

FFurkan Gözükara SECourses welcome

Szutykos Macska•3/23/23, 10:10 PM

No, i thank you!

This is so detailed, dammn

Furkan Gözükara SECourses•3/23/23, 10:26 PM

training started

Furkan Gözükara SECourses•3/23/23, 10:26 PM

Furkan Gözükara SECourses•3/23/23, 10:59 PM

all working good first checkpoint saved. i am not generating any previews . saving checkpoints every 4 epoch

Furkan Gözükara SECourses•3/23/23, 10:59 PM

Zono50•3/23/23, 11:01 PM

I can't figure this out, my training data looks perfect, saved the weights, converted to checkpoint, when i use the prompt in txt2img, it looks nothing like my subject

Zono50•3/23/23, 11:01 PM

my sample data looks perfect but when i use the lora file, it's a completely different person

ZZono50 I can't figure this out, my training data looks perfect, saved the weights, conv...

Furkan Gözükara SECourses•3/23/23, 11:02 PM

you are blending your training into many tokens

Furkan Gözükara SECourses•3/23/23, 11:02 PM

Zono50•3/23/23, 11:02 PM

should i simply do photo of woman instaed?

sample_792-2aa1536b81ae9eda64f837e1587977e79e9327732.png

Zono50•3/23/23, 11:03 PM

sample data

Furkan Gözükara SECourses•3/23/23, 11:03 PM

also

possible to train for style with a couple hundred images? i saw the recent midjourney video uses tho

Similar Threads