Open NeVeREire opened 1 year ago
did you rename the images to one single unknown word ?
First of all, Ben, thank you very much for your work! It's the same for me, I used to train with about 50-60 images and the results were spectacular. I used the values indicated in the colab: UNet 1.500-3.000 , Learning Rate 2e-6, Text Encoder 450. Now I have tried many different combinations and none of them gives good results: from 300 to 900 steps, different Learning Rates, and Text Enconder from 50 to 300. All of them give me this result:
@TheLastBen, what would be the recommended values for training with version 1.5, a person with between 30 and 50 images? Thanks a lot
when resuming training, set the text encoder steps to 0 and use only 10 good images
@TheLastBen, thanks for your quick reply! Forgive my ignorance, but starting from scratch a model, now an ideal configuration would be 10 images, UNet Learning Steps: 300, Net Learning Rate: 2e-6, Text Encoder Training Steps: 0. would be something like this?
from scratch you use the default settings, then when resuming you set the textenc steps to 0
I've tested it twice, once with 30 images once with 15 images, both results show people nothing like the person in the model (which is me this time) .