Arthur C. Doyle's great detective Sherlock Holmes, as depicted in various now public domain covers and illustrations made during Doyle's lifetime. Most of the art is Sidney Paget's, though works by other artists are included. Making a public domain images only character LoRA so I could release the dataset as an example is something I've been wanting to do for a while and finally decided to do it. Couldn't find something I wanted to do with a good quality dataset, so I decided to try a mediocre to poor one and see how it worked. Seems to have come out OK. Works best when style is specified (via prompt or LoRA) instead of left to be generic anime.
Tags: P_D_S_H, black hair, receding hairline
Not in dataset, but in the original fiction: grey eyes
Accessories: smoking pipe,
Not Trained With But Enhances Age In Some Types of Generation: mature male
Common Clothing: coat, vest, collared shirt, cross tie, pants, pointy shoes (outfit varies enough there's no "one" outfit trained, but using these will give a fairly recognizable look. Prepend whatever colors you want)
Secondary clothing pieces: top hat, open jacket, smoking jacket (prepending "silk" can enhance texture here)
Only in dataset once but base model knows fine: deerstalker
Negative: cross (stops crucifix from showing up when "cross tie" is used)
Made with 35 pics. Being primarily cheaply printed black and white images, with the few color images largely being badly decayed book covers, the dataset quality isn’t the greatest but I wanted to try. This required more cleaning than I normally bother with due to the very finite dataset, mostly removing stray heads with some basic GIMP work. If this work badly, I can blame the terrible dataset. Since this data is all public domain I have released the dataset for download (and if my really crude cleaning job, or more likely my tagset, is somehow eligible for copyright, I fully reject it here and release it into the public domain, and CC-0 for where that’s not possible) so you can see what a tagged dataset looks like instead of it being a black box. Used on-site trainer's default settings but with 15 epochs and shuffle tags with keep 1. Since I've given people everything needed to reproduce this LoRA and the concept is very much public domain, I've enabled all permissions for the model.
Also I didn't plan to release this on Halloween night (from the prospective of the Americas), but he's fought fellow public domain character Dracula plenty so...