Who is Claire? Do you recognise her? You don't. She's a creation of Stable Diffusion and is here for your use in your prompts.
I recommend you use ADetailer to show it at its best.
Any buzz tips or ratings appreciated!
Another Faceapp method, trying to do an emo-girl twitch-streamer type, with soft purple hair.
This didn't work at all to start with, so I tried again with the method of just having "woman with purple hair" as the initialization text, something basically none of the tutorials suggest. It worked so well that I went back and tried again on the dataset of Crimson (to see if I could make her head artificially red), Aiko and Venus (to see if I can give them the actual hair color of their data set instead). The answer was no for all of them, which is disappointing - though Crimson's training may be slightly better, so I may re-upload anyway.
I think the problem with Aiko and Venus may be linked to their method of generation - I'll try again with the Faceapp to see if I can get what I was looking for.
Description
FAQ
Comments (3)
I'm really excited about your purple hair discovery. I've been banging the drum about Initialization Text (okay, very softly tapping something drum-like) for a while, but no one's been much interested. I think the problem is that it's not hugely powerful by itself, it's just another tool on your belt. Plus, it's not as useful for creators doing celebrity TIs as it is to those making their own characters. You can "bake in" hairstyle/color, body type, and even some personality traits like "chic," "glamorous," or "shy."
If you tell SD that this TI is a skinny fashionista or an athletic rebel, it will listen but you also need to make sure that you have enough vectors assigned to contain all that information plus the visual stuff. I've been using about 7v in my recent training to hold 20-30 images plus 7-10 keywords. I have no idea whether that's sufficient or overkill. I don't see much difference between 7v and 10v, but I also don't know how to test it or even if I need to train a 10v embed for longer.
Anyway, I appreciate your summaries of what you tried and experimented with on each embed that you publish. Just know that at least one dude finds them super interesting and is reading them all.
This is all really inspiring! I used this in the latest training I did, and it did make a difference to the results - I was trying to create a villain, and with the right prompting, he's definitely got a different kind of vibe. I only did it with my usual 4 vectors, so next time I make a character, I may try with more.
I'm also glad someone finds my notes useful. :)
@Perfection Yes! In most cases it changes the "vibe" of the character rather than something visually obvious, but you can definitely tell. When I was first experimenting with this, I made xyz grids where I just added one keyword to a prompt to see how it affected the image. Some stuff surprised me. For example, "cute," "adorable," and "darling" all made the subject look slightly younger while "pretty" was a softer version of "beautiful" and "gorgeous" and ravishing was the top attractive synonym.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.