Faye is trained from 100% generated images! Any likeness to person(s) living or dead is as coincidental as finding your own doppelganger out there - could happen, not intentional. Her reference images were created using Krea 2, Flux 2 Klein, and Gemini. Then she was put through a bunch of tuning passes.
She is done... Final is a purely re-trained version built from generating about 2,000 images to get proper proportions, angles, shape, and frame out of the multiple failed models - of which, I narrowed to 21 diverse photos to capture as much range as possible. I couldn't leave the problem unfinished. She is complete.
Sick Ollie, Analogue Madness, and anything from Stable Yogi as a base model can drift her age young in some cases, just like her other attempts. The best I can attribute this to is hair, face, and neckline tapping into blocks that are of younger women - she meets the same metrics as Tegan and Carys (and they don't drift young in the same prompts).
Likely will have small lapses of look, as I couldn't get EVERY image to be completely identical to character. Drift you see in the previews is mostly base-model drift (I have tested her in 12 different base models and she mostly carries her look throughout).
For context, I now have about 10 different scripts to extract spatial, textural, and compositional data about characters, 8 different compression and pruning scripts, and 6 re-weighting and mergine/combination scripts. You can't easily fix a poorly trained model! Your work is best suited for building CLEAN and SOLID training data! These tools will (and have already) helped build out more generic LoRA for you all. I hope to release some of them as I get them more broadly tested.
Description
Re-architecture! Used same base training, with quite a few new structural and textural extractions / modifications. She should hold character a LOT better now.
FAQ
Comments (7)
Tempted to archive this one. Not one of my best works, but I have a ton of tools built and will likely be releasing some sliders soon.
I appreciate all the testing and feedback on this lab rat of a model.
No stress. At least we got some good renders so far.
I think you have good taste when it comes to making these characters - i tested lot of interesting / good looking characters, but they often missed the "fun factor" about making them look "alive". And this worked outstandingly well on both Carys, and here on Faye (presuming i got her to look the same between prompt edits).
Well i was looking at the Tegan LORA next, though unsure where to put her artistically. Time will tell. Keep up the good work!
@LurkingAI ha! Fair. Faye has also been great at making tooling around model work.
Appreciate it, I always look for the fun/lively/believable - too many Barbie’s with a choose your hair/face/frame but all are just too same-y.
When you get to Tegan: Glam is nice, but sometimes felt too AI + freckles. Girl Next Door is a bit more fun, has some consistency issues (not as bad as Faye). Balanced should be a solid mix, leaning a bit towards glam look, with some more expression and “normal human” feels.
You have a good range of prompts that stress the characters. Happy Carys has done so well, not only is she my favourite, but it shows a true power in nailing the training.
@LurkingAI as for art style. I think I cast her as the naturally cute, but a bit rough, but not like fit figured - sports, outdoors, cafe/barista/bartender, can rock a good swimsuit look, but also makes lazy wear look good.
Happy to see where you take her though.
I appreciate any and all feedback! I have notes on how I got to this madness of a Final model; I am sure she is still not perfect, but she will generally play much nicer.
Looks good. Currently doing some more shots with Tegan tho!
@LurkingAI Nice :) You can leave Faye for a bit... she's been a real jerk of a model ;) lol