Complete overhaul - entirely new dataset, reg images, settings, etc. Still "good" but not "great"
Some important notes:
Significantly better odds of achieving the hairstyle VS version 1, but less respect towards original image.
Weights needed to achieve hairstyle with minimal image alteration vary wildly on a per-image basis (anywhere from 0.2 to 1.0, even on the same checkpoint + hair color)
Works best on illustrative, realistic, and photographic models
Anime models tend to keep/add the center-forehead bangs regardless of weight
Doesn't work well for Asians (keeps the center-forehead bangs) or Indians (just doesn't apply)
Still some struggles with brunettes (usually just need higher weights for this hair color, e.g. 0.5 to 1.25)
Has a strong desire to significantly alter image/ features, although it tends to adhere to the overall style even so.
You may want to add "" and/or "" to negative prompts
Not tested on dudes. All training data images are of women of various natural hair colors and ancestries.
