V2.0

Two gothic/gothic lolita text inversions

• b_horror_v2_1 is cute and dressy, good for gothic lolita and cute vampires

• b_horror_v2_2 is dark, good for scary vampires

Prompt:

• nsfw

• backlight (all the training samples have a girl dressed in black with a very bright moon or a very bright window)

• colorless (most drawings have low contrast or low saturation)

Other notes:

• this was difficult to train, mostly because all source drawings have low contrast, low saturation, lots of small elements that turn into artifacts, and because most girls have a black corset, the waist area is always excessively dark/shadowed

• correcting brightness/contract, maybe this: https://civitai.com/articles/1125/new-tool-cdcolordetail-tuner

• therefore, a good way to make a drawing without the text inversion is to add "victorian dress, frills, backlight, low contrast, low saturation, 90's style manga" to the prompt

• the model has a tendency to:

  • make girls that hold things in their hands (their dress, cake, flowers, a glass of blood)

  • add some chibi clones (sometimes)

V1.0-Experimental

I was doing sth with the computer, and this appeared.

You need to play with CFG scale, embedding strength, and strength of 'nsfw' in the negative.

Embeddings:

• b_horror_1: vampire, red and black

• b_horror_2: demon, pink and black, scary symmetry

• b_horror_3: nsfw, lots of frills, good for goth lolita

Other notes:

• horror_sth_3 is a mix of 20% horror1, 50% horror2, 30% horror3

• I have included some images with latent upscale, you should try to obtain the same results by hand (it's how intermediate checkpoint examples are generated when training a text inversion)