Use again image caption instead of instance word

[remybonnav]

Hi, Is it possible to use image caption for training? Here you only mention filename that will be the trigger word but you made disappear all options on caption training.

Could you implement it back?

Thanks in advance

[TheLastBen]

this experimental version includes captioning : https://github.com/TheLastBen/fast-stable-diffusion/blob/Captions_dir/fast-DreamBooth.ipynb

[nawnie]

oh my my my what do we have here my friend

[TheLastBen]

The learning rate should be upped a bit when using only captions, I'm yet to test that

[remybonnav]

Yes yes yes. Amazing

On Sun, 18 Dec 2022, 07:55 Ben @.***> wrote:

The learning rate should be upped a bit when using only captions, I'm yet to test that

— Reply to this email directly, view it on GitHub https://github.com/TheLastBen/fast-stable-diffusion/issues/1061#issuecomment-1356704458, or unsubscribe https://github.com/notifications/unsubscribe-auth/AODP7RBASGRHZ3FYJDAXK63WN2YPFANCNFSM6AAAAAATB2IMPA . You are receiving this because you authored the thread.Message ID: @.***>

[Lyntai]

this experimental version includes captioning : https://github.com/TheLastBen/fast-stable-diffusion/blob/Captions_dir/fast-DreamBooth.ipynb

Nice. If I want to train a specific person or face, should I input the token in the caption text or just "a person..."/"a face..."?

[TheLastBen]

if you caption the images you should include a unique identifier (that replaces the name) of person in all the images captions. But I recommend not using captions for training a face, use higher learning rate instead