The model is trained with clip skip 1 on top of ACertainty
The locon is extracted from the full model
The original one seems to give worse result when applied to certain base model so I updated with one with smaller compression ratio
The associate hugging face repository contains other versions of models https://huggingface.co/alea31415/watayuri
Trigger words:
ChibanaSumika
ChibanaSumika, glasses
KoshibaMai
MamiyaKanoko, glasses
MamiyaKanoko, hair bun
NishideraNene
ShirakiHime
ShirakiHime child
YanoMitsuki
YanoMitsuki child
There are very few images of Nene and Sumika without glasses for now so we probably cannot get any reasonably good results.
Moreover, the model is only trained with screenshots for the moment, so all the images are tagged with aniscreen. You can put it in either positive or negative (putting it in negative could be better when switching to some base models).
Characters are separated by ;.
You can get the clothes correctly with general prompts like green school uniform or brown coat.
--
I don't need support or credit, but I would be glad to know that you are using the models I trained and find it useful.
Moreover, I would like to advocate for more franchise models.
You can take a look at my workflow https://github.com/cyber-meow/anime_screenshot_pipeline if you are interested.
I just want to spread the fact that there is no reason to encode a single concept in each lora.
Description
FAQ
Comments (5)
The effect is particularly good！
I'm curious about how you train multiple characters in a training set, Lycoris parameters, that sort of thing. Can you tell me?
I've trained a multi-role Lora before, but it can't be two people at once...
For the training set you can take a look at the gihtub link https://github.com/cyber-meow/anime_screenshot_pipeline
In brief I just use all the screenshots that I can get from the anime (including both single- and multi-character frames, the key to be able to generate two people at once is to have such images in training set)
For LyCORIS I have been sticking to lr 2e-4 for a while. Most of the time I use batch 8 for around 30K steps. However I have a also a regularization set of 30K images that is used around half of time, so in total it is like around 120K uses of relevant images. For LoHa I use dim 8 alpha 4, conv dim 4, conv alpha 1.
I have not tried to optimize too much these things, so many other configurations could work as well. For example it is often suggested to use lr 5e-4, for which the training may be faster.
So glad to find yet another person making girlpacks. Thanks for all the hard work!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.