Type | |
Stats | 3,727 |
Reviews | |
Published | Apr 22, 2023 |
Base Model | |
Training | Steps: 4,960 Epochs: 32 |
Trigger Words | star butterfly green squid outfit green spider outfit horn hairband shroom hairband shroom outfit rainbow outfit nightgown sleep mask princess dress |
Hash | AutoV2 EF939891BC |
If you like the LoRa, please feel free to share your generated images here and leave a review (or at least a rating) for the sake of motivation and stuff ;D
For usage guidance, look into the About this version box for more information and click on Show more to get all trigger words and combinations (The trigger words in the Details box are a pain to manage).
Please contact me here on civitAi or on Twitter if you want to provide the LoRa in a generation service.
About
Well, this week I realized that there are more than 2 sessions of this series. (I watched it years ago, and it was nice back then ^^).
So I went back and watched the first two sessions, and occasionally made screen caps of the characters, because I was thinking: I want to make LoRa available to all the characters that I care about.
I start the first batch with Star Butterfly the magic crazy girl, in most of her outfits (she changes them quite often, which is nice, considering that this is a western animation series). And the existing LoRa didn't tag the outfits properly, which is the main reason why I created my own ^^.
Additionally there will be Marco's Mom and Jacky Lynn coming up this weekend, they are currently in training :D
Training
This was fun and due to the fact I wanted to train most of her outfits (there maybe some missing) I ended with 155 images. (v1 had just 128 but I actually forgot to put two more outfit to the training data. So I created v1.1)
Additionally instead of using repeats I moved to use epochs instead... this way I have way more control over the end result ^^ (But I think I nailed it with 32 epochs ;D)
This is the first time I've specifically tagged the model for specific outfits, it worked quite well in most cases (first I used deepbooru for tagging, but there was just too much unnecessary junk, so I started from scratch, and although the dataset tag editor helped a lot, it was still a pain).
Maybe I'll make an updated version to get a bit more training data for the outfits that didn't go so well. I only used the animated series, but maybe fanart / the comics could solve the problem of those missing outfits ^^