fast-stable-diffusion icon indicating copy to clipboard operation
fast-stable-diffusion copied to clipboard

Caption blending,

Open nawnie opened this issue 2 years ago • 4 comments

while adding 2 different people, (faces:NO text encoder at 85% steps are 800 each{cant believe im getting good ones as low as 500} roughly 25 images for each)

the two people begin to blend, the newest model is almost correct but any previous model seems to take aspects of the new one. Names used for testing us xyz and abc, real names where used when noticing this.

maybe merging would yeild better results then adding new models to the project?

nawnie avatar Nov 12 '22 08:11 nawnie

if the two people are from the same gender, they will merge, you can reduce that effect by adding more people, but also more steps.

TheLastBen avatar Nov 12 '22 09:11 TheLastBen

interesting, it does seem that adding 2 at the same time works just fine. just doing one then the other causes some issues. ill try at higher steps later.

nawnie avatar Nov 12 '22 09:11 nawnie

if you caption on "person", then the other also "person", it's only logical that it will merge them, that's why I recommend avoiding captioning the images

TheLastBen avatar Nov 12 '22 10:11 TheLastBen

I used the term Caption due to its reference in code being Caption = file name

But as I said the only Caption ie file name was just ABC and XYZ .

Wheb we talked previously about using tags that was for a speacil project with 1 subject, worked well. But yea sadly can't seem to use tags with multi subjects, unless you tag them with diffrent names, I've had success using younger for one subject then regnouy for subject 2, but in the end you where right, they are more of a pain then helpful.

sd model 2.0 needs to have a separate storage space for tags like some are using for class and instance, a third for tags that would muddy the instance woukd be nice.

Anyways ty again man.

On Sat, Nov 12, 2022, 4:22 AM Ben @.***> wrote:

if you caption on "person", then the other also "person", it's only logical that it will merge them, that's why I recommend avoiding captioning the images

— Reply to this email directly, view it on GitHub https://github.com/TheLastBen/fast-stable-diffusion/issues/464#issuecomment-1312447769, or unsubscribe https://github.com/notifications/unsubscribe-auth/AZPYLSELUJWJFS3O7TRBVRDWH5VW5ANCNFSM6AAAAAAR6G2ZPI . You are receiving this because you authored the thread.Message ID: @.***>

nawnie avatar Nov 12 '22 17:11 nawnie