fast-stable-diffusion
fast-stable-diffusion copied to clipboard
TheLastBen
Pls help:training multiple people with 220pics, how much text encoder should I set?
hi, Ben: I want to trained 7person, each person 30~35 imgs. (I knew it may has blend problem...) how much text encoder steps in total should I set? what's the best percentage?
I begin with 6600steps+1050 text encoder at first training, when I resume, I'm not sure whether or not should I still set text encoder training steps... (I resumed 6600 steps each time)
thx for you time.
15 -20% of total steps, SET LR to 1e-6 to attempt to reduce blending 7 people is definitely doable. its going to take a few extra steps then normal. if you wish to make use of concept images add more images of the same people to that folder, it will help, train that one 2x as much as normal text
reduce people to 20 to 25 a piece, more photos takes alot more steps, each change in face or diffrence in time picture was taken (subject hair length weight etc) will make the model take longer to train, if your doing multiple people try to find the less amount of photos possible.
dont train Text_encoder past 1k steps no more then 600 on your first run is what i follow. if a subject has a unique trait add a tag for it ie xxx is the only subject with brown hair, title xxx_brown_hair, dont add similar tags only tag things unique to the subject, each tag will act as a token enforcing that subjects difference from the others. let the text encoder find the class. do not tag man woman people. that's just asking for blending issues.
when prompting your subject if at first you dont succeed start adding the tags you used in your filenames, each tag should bring you closer to your target, this is important due to some targets taking longer then others to train you may pull xxx fine by there name but need to add yyy_Tag1 for that one.
this will fry your models ability to use these tags with other things, such as emma stone with brown hair, will now resemble your subject.
im assuming this is a personal model where cutting corners is ok since your not asking how to train multiple subjects and a style, cause that's opening another can of worms. and that has an answer on several post now too if you search, wish ya luck bud
15 -20% of total steps, SET LR to 1e-6 to attempt to reduce blending 7 people is definitely doable. its going to take a few extra steps then normal. if you wish to make use of concept images add more images of the same people to that folder, it will help, train that one 2x as much as normal text
thank you so much. nawnie. what's "SET LR to 1e-6 "? where can I set it? and how can I "add tag for a img"? I haven't find it in fast-DreamBooth colab...
It's how you name the file as in the example. Sadly adjusting lr in this repo requires you to edit some code.
On Thu, Dec 15, 2022, 9:21 PM yasmineyxz @.***> wrote:
15 -20% of total steps, SET LR to 1e-6 to attempt to reduce blending 7 people is definitely doable. its going to take a few extra steps then normal. if you wish to make use of concept images add more images of the same people to that folder, it will help, train that one 2x as much as normal text
thank you so much. nawnie. what's "SET LR to 1e-6 "? where can I set it? and how can I "add tag for a img"? I haven't find it in fast-DreamBooth colab...
— Reply to this email directly, view it on GitHub https://github.com/TheLastBen/fast-stable-diffusion/issues/1031#issuecomment-1354137389, or unsubscribe https://github.com/notifications/unsubscribe-auth/AZPYLSHOUFCQRH5KDKPMKXLWNPN2NANCNFSM6AAAAAAS7MMROY . You are receiving this because you commented.Message ID: @.***>
It's how you name the file as in the example. Sadly adjusting lr in this repo requires you to edit some code. … On Thu, Dec 15, 2022, 9:21 PM yasmineyxz @.> wrote: 15 -20% of total steps, SET LR to 1e-6 to attempt to reduce blending 7 people is definitely doable. its going to take a few extra steps then normal. if you wish to make use of concept images add more images of the same people to that folder, it will help, train that one 2x as much as normal text thank you so much. nawnie. what's "SET LR to 1e-6 "? where can I set it? and how can I "add tag for a img"? I haven't find it in fast-DreamBooth colab... — Reply to this email directly, view it on GitHub <#1031 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AZPYLSHOUFCQRH5KDKPMKXLWNPN2NANCNFSM6AAAAAAS7MMROY . You are receiving this because you commented.Message ID: @.>
thx for u reply. I didn't find the example to rename the file. (in the previous edition, it did had example, but not mentioned how to add tag）. I now just rename file in this format(eg.hma=person name): hma(1),hma(2).. if I want to add tag, what's format should I follow? Is "hma(1)_brown_hair" right? or"hma_brown_hair(1)"?
You got it
On Thu, Dec 15, 2022, 10:34 PM yasmineyxz @.***> wrote:
It's how you name the file as in the example. Sadly adjusting lr in this repo requires you to edit some code. … <#m_-897023892436360122_> On Thu, Dec 15, 2022, 9:21 PM yasmineyxz @.> wrote: 15 -20% of total steps, SET LR to 1e-6 to attempt to reduce blending 7 people is definitely doable. its going to take a few extra steps then normal. if you wish to make use of concept images add more images of the same people to that folder, it will help, train that one 2x as much as normal text thank you so much. nawnie. what's "SET LR to 1e-6 "? where can I set it? and how can I "add tag for a img"? I haven't find it in fast-DreamBooth colab... — Reply to this email directly, view it on GitHub <#1031 (comment) https://github.com/TheLastBen/fast-stable-diffusion/issues/1031#issuecomment-1354137389>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AZPYLSHOUFCQRH5KDKPMKXLWNPN2NANCNFSM6AAAAAAS7MMROY https://github.com/notifications/unsubscribe-auth/AZPYLSHOUFCQRH5KDKPMKXLWNPN2NANCNFSM6AAAAAAS7MMROY . You are receiving this because you commented.Message ID: @.>
thx for u reply. I didn't find the example to rename the file. (in the previous edition, it did had example, but not mentioned how to add tag）. I now just rename file in this format(eg.hma=person name): hma(1),hma(2).. if I want to add tag, what's format should I follow? Is "hma(1)_brown_hair" right? or"hma_brown_hair(1)"?
— Reply to this email directly, view it on GitHub https://github.com/TheLastBen/fast-stable-diffusion/issues/1031#issuecomment-1354191359, or unsubscribe https://github.com/notifications/unsubscribe-auth/AZPYLSFX65BHK4AQVZD3NJLWNPWMVANCNFSM6AAAAAAS7MMROY . You are receiving this because you commented.Message ID: @.***>