1st try at a LORA. Used the kohya LORA collab to make this Camilla LORA.

The different versions are from different epochs. I personally recommend 03, but I have noticed 04 can work better on more realistic / semirealistic models.

I recommend keeping the strength below 0.8, especially if you use the 04 version. It's gonna vary from model to model, and whether you are trying to get an image that is similar to Camilla or extremely similar.

Known issues:

• Hands have a tendency to be raised and holding her hair or have 1 hand up and another down

• Eye color doesnt always come out purple

• Her horn crown shows up usually at higher values if you don't include fake horns in the negative

• Back views are kind of sus because no training pics were from behind

• Some models / LORAs you may need to specify a hair over one eye tag for consistency (especially face LORAs, like an ahegao one)

• I've found if you don't add things to the prompt about her curves (eg: curvy, cleavage, breasts, large breasts, thick thighs, etc) in some models you get very un-camilla looking results

eg: adding just 2 descriptors about Camilla radically changed this result

Training data info:

• Sample consisted of 52 images

• tags were set manually

• Model used to train was anything v3.0 w/ the anime.vae

Other embeddings / models / LORA etc used in example imgs:

• nerfgun - bad prompt

• yesyeahvh - bad hands 5

• Ulzzang-6500

• bloodsuga - perfect world (baked vae)

• anything v3

• anything v4.5