Replies: 3 comments 2 replies
-
If pointing at a directory of existing images, the app will just use that
directory. But, class images now need to match the resolution of their base
input images. Without any other information, my best guess is you maybe
generated class images before bucketing was implemented, and then I fixed
bucketing, which makes the requirement change.
…On Sun, Feb 5, 2023 at 5:53 AM Apocangel ***@***.***> wrote:
tl;dr I have a set of classification images already generated by DB, but
every time I try to train, it starts creating new ones.
Dataset is unchanged. Settings are the same.
Any way to force DB to use existing ones? Dataset is unchanged. Settings
are the same.
The long story:
I'm on latest on all.
running on Auto1111 SD
I am training a LoRA model with a dataset of about 1000 images, and need
to create about 5 classification images. Thats 5k Classification images to
generate. I pre-generated them all which took about 40 hours with my
hardware. Then I tried starting the training, but it started generating new
classification images all over again in the same folder it created them
previously (./classifiers_0/). I let it run and after another 40 hours, it
started training. I had to stop to update SD. After restart, it started
generating new classification images a third time.
I tried removing some images from dataset, and it still tries to
regenerate new ones, albeit a lower number. I tried removing some of
dataset to force it to regenerate only the missing ones, but it tries to
regenerate the whole set.
Any way to force DB to use existing ones?
—
Reply to this email directly, view it on GitHub
<#899>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAMO4NBFUN4RBVQXJLI74RTWV6IDHANCNFSM6AAAAAAURXKRZY>
.
You are receiving this because you are subscribed to this thread.Message
ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
Nope. Nothing has fixed the problem so far. |
Beta Was this translation helpful? Give feedback.
-
It's basically what D8 said. The class images are linked to the instance images. There have been some bugs related to how to it found them in the past. However, on the latest main commit, class images should be working properly. If you use class images you need to set the # of class images per instance for each concept in the UI. for example if you set it to 1, it will try to find 1 class image that matches each instance pic. The resolutions must match the post-bucket Ed instance pictures. If you are using captions, the captions must match too. The extension will create the class images that it needs and can't find. |
Beta Was this translation helpful? Give feedback.
-
tl;dr I have a set of classification images already generated by DB, but every time I try to train, it starts creating new ones.
Dataset is unchanged. Settings are the same.
Any way to force DB to use existing ones?
The long story:
I'm on latest on all.
running on Auto1111 SD
I am training a LoRA model with a dataset of about 1000 images, and need to create about 5 classification images. Thats 5k Classification images to generate. I pre-generated them all which took about 40 hours with my hardware. Then I tried starting the training, but it started generating new classification images all over again in the same folder it created them previously (./classifiers_0/). I let it run and after another 40 hours, it started training. I had to stop to update SD. After restart, it started generating new classification images a third time.
I tried removing some images from dataset, and it still tries to regenerate new ones, albeit a lower number. I tried removing some of dataset to force it to regenerate only the missing ones, but it tries to regenerate the whole set.
Any way to force DB to use existing ones?
Beta Was this translation helpful? Give feedback.
All reactions