Coming from JoePenna / Dreambooth-Stable-Diffusion

[metzo007]

Hi,

I started Dreambooth using the JoePenna Dreambooth Repo. I always used the person token, and roundabout 20 pictures of the same person ( 5 of the face, 7 of the upper body and 8 of the whole body), awoiding to have more than 2 with the same backgound or light, and 2000 steps.

I changed to fast-dreambooth using the same settings. 20 pictures with the same random name and 2000 steps gave me similar results as JoePenna Dreambooth before.

But now with captions, concept images, 3 different trainings methods with separate learning rate settings and the reduction from 20 to 10 training images as starting point a lot has changed.

I did leave everything by default, and only increased the UNet_Training_Steps from 650 to 1300 to account for using 20 instead of 10 images. I got ok results in overall quality of the picture, but the model is very unresponsive to any prompt.

Can someone share good settings coming from JoePenna? I thing for training a person captions are not to usefull but concept images might help.

@TheLastBen. I think a easy-dreambooth with only a photo upload is a good idea. :-)

Best, metzo

[TheLastBen]

Hi, if you want to train on a person, you don't need captions or concept images, keep everything to default, only resume the unet training for 100 steps at a time after testing. the V2 also give better results. make sure you're using a unknown identifier as the images filename without adding (person) or (man/woman)

[TheLastBen]

8e-6 is fine

[metzo007]

That sparks my curiosity. :-) what is the difference between a lower learning rate and lesser steps?

[TheLastBen]

let's just say that the learning rate is a more aggressive method for a faster training and also a very low learning rate will make the training useless, the number of steps are effective only when the learning rate is correctly set.

[ynlissaspill]

V2 also give

on dinos the same like 1.5

[metzo007]

dinos?