i'm using a transparent channel on my subject tho so i lowered the learning rate significantly and c
i'm using a transparent channel on my subject tho so i lowered the learning rate significantly and checked "Use PNG Alpha channel as loss weight"
Fi think you need extra processing for this. it was like that initially "Use PNG Alpha channel as loss weight"
Fi mean pre processing of images
Fbut i dont know if they automated it
Zyeah i used the extension background remover for it and did alpha matting
ZGitHub
Removes backgrounds from pictures. Extension for webui. - GitHub - AUTOMATIC1111/stable-diffusion-webui-rembg: Removes backgrounds from pictures. Extension for webui.
Zwhy try to do captioning and file words and all that when I can completely eliminate the background and focus on the subject which is what i want anyways
Zi still use a photo of a [name] tho
Zshe's looking decent at 1500 steps already, she has very unique eyes so will probably have to train much longer than usual
Fye i saw this it is a good one
Fmuch more logical
Feyes are harder
Fi recorded a new dreambooth install video because people fails 
CHi man. I just added this extension. How do i use it? Sorry i am new to this
Zit's under the extras tab
Zyou 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
Zfor the embedding inspector, i know you want loss to stay under 0.2, but what about magnitude?
Zor is it different for every data set?
Fit is different for every dataset. best way is checkpoint comparison
Zdoing the xyz plot, is there a way to do multiple textual inversions as well as cfg?
Fyou want to use more than 1 in 1 run?
Fyes you can do with s/r prompt
Zwell what's the best way to compare multiple textual inversions
Zcause i see a couple that have images that are really close
Fyou want to compare checkpoints right?
Znot checkpoints, textual inversions
Ftextual inversion savings
Zembeddings
Fso we can say as checkpoint
Fok they are activated with filename
Fso you can use s/r prompt
Fwhich is looking for a keyword
Fand replaces it
Fi explained it in my videos
Zoh right, so photo of [embedding[, then replace would be the different embeddings
Zgot it, thanks doc
CDo you know whats the best one to choose from the selection for remove background?
Zyou just have to play around with it, turning on return mask and trying each one will give you a visual of what they do
C@Zono50 thank you for the help
Zabsolutely
COne more question, how low did you go with your learning rate. ive seen a reddit post saying to reduce by 30%?
Ci normally use the default
Zyeah 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
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
Zwhich 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
Zwell actually at 7500 steps, she's looking pretty good
Zit'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.
ZWhat do you think? pic on the left is the actual girl
Zwomen always think they look different than they actually do, it's their perception of themselves. I think it's pretty spot on
2is it ok to ask question here about errors in a google colab? or is there a more appropriate chanel?
2I'm getting an error when trying to train the model, unrecognized arguments: portraits