My first attempt at making a character LoRA (It didn't magically turn out great ^^): Brigette Lundy-Paine, now Jack Haven, probably best known from their role as Casey Gardner in the TV series Atypical. (Since this focuses on them before their coming out as quite masculine-leaning non-binary, I will be referring to them as "her" below)This was trained on 145 images of middling quality, mostly headshots, red carpet appearances and screenshots from movies and TV that I just gathered from image search.It has a tendency to produce flash photo lighting, lots of makeup, shiny skin, sometimes rather grainy images, and quite tightly cropped headshots (one of the most frequent problems I have with SD 1.5 anyway) by itself, so it's best used with mitigations like<pre><code>grainy, low quality, (black/white, monochrome, greyscale:0.3)
(closeup, zoomed in, cropped:1.3), (makeup:1.1), (flash photo:1.3), long hair</code></pre>in your negative prompt, depending on your needs, you know the deal; and supplementing LoRAs.I used the two CoLab spaces from <a rel="ugc" href="https://civarchive.com/models/22530/guide-make-your-own-loras-easy-and-free">this guide</a>, with mostly default settings.It's not a great result. I spent probably just over an hour collecting random images for her name as the search term; in hindsight, I should've probably also searched for names of roles she's played; if nothing else to offset the distinct red carpet look of flash photography and excessive makeup; Then I spent less than an hour hand-labeling some pictures before I handed off the rest to the dataset maker CoLab and had CLIP fill in the rest.But it's something, and might be fun tinkering around with, like adding it into a prompt more as a spice rather than as the main ingredient.Sometimes she looks like a thumb, sometimes SD can't decide whether to give her teeth or just gums, you know, the usual 😅If you can spot telltale signs of something I should have done differently from your own experience, please let me know!I'm most interested in any feedback for this (you know, constructive feedback. I know it's not great and could've done with a better dataset, no need to remind me of that 😬)It's unfeasible for me to run SDXL, let alone Flux on my potato, so I'm sticking with SD1.5

Files with SHA256: e4f354f5c7d70ee36ae1d474b19c74cc35ab4e0108fe77ac38bdd7c705e6eb1f