thanks for the insights!

Enjoyed the video very much! I think you're missing out on the Initialization Text. It doesn't make a huge impact, but it does alter some things especially with all those vectors looking for data. Think about it this way: If you type nothing, you're telling the AI to have no preconceived notions; if you just type "woman," it's slightly more informative but makes no real difference; now, what if you type "athletic, tanned girl from Brazil with a lithe musculature and a passion for beach volleyball"?

As a fellow nerd, I've had a good time trying different adjectives in xyz grids to see how they change images and by how much. My non-scientific tests so far suggest some especially loaded words include: "passion, lithe, athletic, lean, shy, adventurous, and ravishing." I don't want to oversell the effect. It's subtle. However, I've been able to nudge embeddings in the directions I want just by using Initialization Text. It doesn't always work the way I think it should, though. I'm not sure if it's an untapped resource or a neat trick that will entertain consistent character creators way more than it deserves.