v3.1 (non-TE)
Updated to base model to Anima Preview 3.
Also tried to split dataset into high and low quality images, so it would initially train on all images for like 4-8 epochs and then only train on high quality ones at the end.
This one might be worse than v3.0 because of this. Womp womp.
v3.0 (non-TE)
I have been trying all week tinkering with this new version to try and include more concepts, and I think it ended up quite good, finally. I think this may be the last OMORI LoRA I'll produce, as I'm satisfied enough with how it ended up.
I had to hack sd-script's prompt processing because I wasn't satisfied with how limiting it was. So I rewrote the function for that to make wildcards, tag dropout and tag shuffling more robust. Unfortunately, though, I don't think it helped much with making the model remember all of the concepts.
Here's a list of concepts that it supports outside of Anima's default capabilities, as well as list of additional trigger words that may help:
something (omori), dark body, one eye
sweetheart (omori), pink hair, twintails, pink dress, hole in stomach (only when her stomach area is visible, like when she has swimwear on)
mutantheart (omori), asymmetrical eyes, slime, pink hair, purple eyes (this one is probably pretty bad)
stranger (omori), black skin, black hair, white eyes, dark clothes
kim (omori), brown hair, blue hoodie, glasses, orange eyes (idk which eyes are good, maybe orange, brown or yellow?)
polly (omori), brown hair, white dress, bare shoulders, brown eyes
hero (faraway) (omori), brown hair, brown eyes, blue sweater, small white emblem
I've found that non-TE version is better this time and it beats out the one with TE trained.
v2.0 (w/ TE and without)
I have recently found out that my training script was also training Text Encoder by default. So I've decided to try training both with and without TE. The file size for LoRA without training TE is obviously smaller because it doesn't have to include the additional weights for TE, but I also found it worse and preserving the style, while better at just the composition.
The newer version just means that it:
1. Has more images in it's dataset.
2. Has slightly more concepts.
Here are some things that are now supported by this LoRA:
Hero actually has the small white emblem on his sweater
Something (that dark ghost-like thing with one eye)
Sweetheart
Better (probably) realistic background.
I recommend using photo background for realistic background as that's what I used in the dataset, replacing real world background and 3d background .
I will try to add support for stranger (omori) and maybe polly (omori) and kim (omori) later on.
v1.0 (styles)
This lora was trained on like 66 images. Anima is a very good base model! I was able to generate another character lora in just 10 minutes (I will upload one after this).
To get the OMORI style, you must add something like @omocat or/and omocat (style)
This LoRA can also do realistic backgrounds using these tags:real world background and 3d background (the latter is usually not needed)
I've also noticed that this version likes to sometimes put black borders, so put that in the negative prompt.