Type | |
Stats | 406 |
Reviews | (50) |
Published | May 28, 2023 |
Base Model | |
Training | Epochs: 8 |
Trigger Words | amira antennae dark skin cat ears |
Hash | AutoV2 98CA453028 |
Version 5:
Version 5 epoch 8
This version was trained less, which means that there is a lower chance that the resulting image gets deep fried. But it also means the character becomes more human and less Aisha.
I also added the antennae tag in the training data, so the antennae (Aisha's extra ears) is removable if it is put in the negative prompt.
The outfits are also more adjustable, the previous version sometimes forces Amira to wear her default outfit.
For some reason she has an animal nose now, I tagged it in the training dataset but putting animal nose in negative prompt doesn't seem to do much.
Version 2:
The previous version had problems with giving random blue faces and very thick outlines.
I think it was because of the tagging, there were a couple of images with a blue dog as a pet that wasn't tagged. This version fixes that. I also added 14 more images which were good generations from the previous version with some touchups to remove the errors. So this version was trained on 34 images instead of just 21 in version 1.
10 repeats, 18th epoch was selected.
Version 1:
Not too sure about this one because I only got 21 images to train from and some are really low resolution. Because if that I did 15 repeats, so in 20 epochs there were 3000+ steps.
If there are enough good images generated from this version (with different poses and backgrounds) I will be willing to make a version 2.
Epoch 12 and 14 gave good results depending on the base model used, in the end I picked 14 even though there are some serious errors for some models, though working on the prompt might also circumvent this.