Software Engineering Courses (SECourses)•3y ago

she's looking decent at 1500 steps already, she has very unique eyes so will probably have to train

she's looking decent at 1500 steps already, she has very unique eyes so will probably have to train much longer than usual

ZZono50 this extension > https://github.com/AUTOMATIC1111/stable-diffusion-webui-rembg

Furkan Gözükara SECourses•3/17/23, 12:00 PM

ye i saw this it is a good one

ZZono50 why try to do captioning and file words and all that when I can completely elimi...

Furkan Gözükara SECourses•3/17/23, 12:00 PM

much more logical

ZZono50 she's looking decent at 1500 steps already, she has very unique eyes so will pro...

Furkan Gözükara SECourses•3/17/23, 12:00 PM

eyes are harder

Furkan Gözükara SECourses•3/17/23, 12:05 PM

i recorded a new dreambooth install video because people fails

ZZono50 this extension > https://github.com/AUTOMATIC1111/stable-diffusion-webui-rembg

CPI_MrRagingBull 🇬🇧•3/17/23, 12:49 PM

Hi man. I just added this extension. How do i use it? Sorry i am new to this

Zono50OP•3/17/23, 12:51 PM

it's under the extras tab

Zono50OP•3/17/23, 12:52 PM

you can use any of the remove background types, return mask will give you an alpha mask, alpha matting allows you to customize how close it gets to the outline

Zono50OP•3/17/23, 12:54 PM

for the embedding inspector, i know you want loss to stay under 0.2, but what about magnitude?

Zono50OP•3/17/23, 12:54 PM

or is it different for every data set?

ZZono50 or is it different for every data set?

Furkan Gözükara SECourses•3/17/23, 12:57 PM

it is different for every dataset. best way is checkpoint comparison

Zono50OP•3/17/23, 1:06 PM

doing the xyz plot, is there a way to do multiple textual inversions as well as cfg?

ZZono50 doing the xyz plot, is there a way to do multiple textual inversions as well as ...

Furkan Gözükara SECourses•3/17/23, 1:07 PM

you want to use more than 1 in 1 run?

Furkan Gözükara SECourses•3/17/23, 1:07 PM

yes you can do with s/r prompt

Zono50OP•3/17/23, 1:07 PM

well what's the best way to compare multiple textual inversions

Zono50OP•3/17/23, 1:07 PM

cause i see a couple that have images that are really close

Furkan Gözükara SECourses•3/17/23, 1:08 PM

you want to compare checkpoints right?

Zono50OP•3/17/23, 1:08 PM

not checkpoints, textual inversions

Furkan Gözükara SECourses•3/17/23, 1:08 PM

textual inversion savings

Zono50OP•3/17/23, 1:08 PM

embeddings

Furkan Gözükara SECourses•3/17/23, 1:08 PM

so we can say as checkpoint

Furkan Gözükara SECourses•3/17/23, 1:08 PM

ok they are activated with filename

Furkan Gözükara SECourses•3/17/23, 1:09 PM

so you can use s/r prompt

Furkan Gözükara SECourses•3/17/23, 1:09 PM

which is looking for a keyword

Furkan Gözükara SECourses•3/17/23, 1:09 PM

and replaces it

Furkan Gözükara SECourses•3/17/23, 1:09 PM

i explained it in my videos

Zono50OP•3/17/23, 1:09 PM

oh right, so photo of [embedding[, then replace would be the different embeddings

ZZono50 oh right, so photo of [embedding[, then replace would be the different embedding...

Furkan Gözükara SECourses•3/17/23, 1:09 PM

yes

Zono50OP•3/17/23, 1:09 PM

got it, thanks doc

ZZono50 it's under the extras tab

CPI_MrRagingBull 🇬🇧•3/17/23, 1:09 PM

Do you know whats the best one to choose from the selection for remove background?

Zono50OP•3/17/23, 1:12 PM

you just have to play around with it, turning on return mask and trying each one will give you a visual of what they do

CPI_MrRagingBull 🇬🇧•3/17/23, 1:14 PM

@Zono50 thank you for the help

Zono50OP•3/17/23, 1:14 PM

absolutely

ZZono50 i'm using a transparent channel on my subject tho so i lowered the learning rate...

CPI_MrRagingBull 🇬🇧•3/17/23, 1:15 PM

One more question, how low did you go with your learning rate. ive seen a reddit post saying to reduce by 30%?

CPI_MrRagingBull 🇬🇧•3/17/23, 1:15 PM

i normally use the default

Zono50OP•3/17/23, 1:15 PM

yeah right now i'm doing 0.005:200, 0.002:500, 0.001:3000, 0.0005
but 30% less would be 0.035:7, 0.014:14, 0.007:42, 0.0035:140, 0.0014:350, 0.0007:2100, 0.00035

Zono50OP•3/17/23, 1:16 PM

which this subject kinda turned out ok that i'm working on around 5500 steps, but now she is looking kinda deformed so may have to stop it here soon

Zono50OP•3/17/23, 1:17 PM

well actually at 7500 steps, she's looking pretty good

Zono50OP•3/17/23, 1:17 PM

it's hard cause sometimes you will get results tha tlook like this and you think it's time to stop, but then the next result looks like the one where she's in blue.

Zono50OP•3/17/23, 1:19 PM

What do you think? pic on the left is the actual girl

Zono50OP•3/17/23, 1:20 PM

women always think they look different than they actually do, it's their perception of themselves. I think it's pretty spot on

2020Throw2020Away•3/17/23, 1:21 PM

is it ok to ask question here about errors in a google colab? or is there a more appropriate chanel?

2020Throw2020Away•3/17/23, 1:23 PM

I'm getting an error when trying to train the model, unrecognized arguments: portraits

ZZono50 What do you think? pic on the left is the actual girl

CPI_MrRagingBull 🇬🇧•3/17/23, 1:24 PM

looks almost spot on. as you say the eyes are the tricky one but looks good

Zono50OP•3/17/23, 1:30 PM

thank you

ZZono50 thank you

CPI_MrRagingBull 🇬🇧•3/17/23, 1:31 PM

Promise last question. Once you remove backgrounds do you blip your new dataset?

Zono50OP•3/17/23, 1:34 PM

it wouldn't let me

Zono50OP•3/17/23, 1:34 PM

like it kept throwing errors, i don'tthink it likes it, but you shouldn't have to caption cause your subject is the only thing in the picture

CPI_MrRagingBull 🇬🇧•3/17/23, 1:35 PM

thank you going to now give it a go with new data set

Zono50OP•3/17/23, 1:36 PM

Great! can't wait to see the results. I posted some scary stuff yesterday

EssaBalls•3/17/23, 1:45 PM

For past two days I was working with Lora and I could do training without problem. All of the sudden I can't train anymore because of the CUDA out of memory...
Is there any way to lower the vram usage for dreambooth / lora training?

she's looking decent at 1500 steps already, she has very unique eyes so will probably have to train

Similar Threads

Similar Threads

Similar Threads