Illustrious is up
Changes in V2: Sorry for the delay- I've been bouncing back and forth between Val and a few other LoRAs that have been giving me a hard time. Many days and 23(!) not-quite-there iterations later I finally have a V2 I'm satisfied with posting. Gens seem a lot more consistent now- the bodysuit seems to occasionally drop a thigh line or the second, thinner arm line, but it's a lot better. Her physical appearance still varies but that's a problem with the source material. Not much I can do there unfortunately. Finally, if it's ever relevant that you want her feet to be visible without the bodysuit simply becoming "skin tight" to her feet, prompt white boots. Alright, time to continue to beat my head against the wall with these other LoRAs.
Valeria Richards in her sublime Future Foundation suit from the comic panel responsible for a million uohhhhs 😭😭😭. If you've seen it, you know what I'm talking about. Side profile, hair twirl, red guy. Great stuff. Despite being drawn by what seems to be a different artist in each panel in the dataset, she generates pretty well!
Sometimes she generates without her bangs; just throw bangs in your prompt if it keeps happening. Same with blue eyes if her eyes aren't blue. But I wouldn't do this unless you have to because it seems to make the outputs more... generic? I don't really know how to describe it. And if you end up liking the look, by all means keep the corrective tokens in.
For V2 I plan on going over each image in the dataset and removing the backgrounds in the hopes that it gets rid of the overfit on rocks/rock textures and the color orange, since those seem to be overpowering the generic backgrounds I'm trying to gen.
Enjoy!
Description
FAQ
Comments (2)
Did you use "bangs" in the training data tagging? If not, that's likely why it makes the output more generic: It's just the checkpoint forcing its base idea of "bangs" onto the image.
As for the backgrounds: I think just tagging the rocks and "orange background" in training data (if you didn't already) would keep those in line since that works for artstyle things like 3D. Would certainly be more pleasant than full background removal (easily the biggest time crunch and most tedious step in data cleaning, which I only do if I absolutely have to. But I'm also very unskilled at GIMP.)
Edit: A big issue appears to be a lack of full body images. It has no idea how to do below the calves. May be source material weakness though.
I didn't tag bangs, but I did just realize there's a single image of her with her hair up in the front that I neglected to tag- definitely the culprit there. And good point on the background tagging, that shit takes me hours. You're correct in assuming that full body shots are sparse- there's a single one in the dataset currently. I'm gonna go back and see if there are any clear shots of others wearing the suit that I can just crop the heads from. Appreciate your help!