It's pronounced KEER-uh :)
I was inspired (by @RaymondLuxuryYacht of course!) to try making a character LoRA. I don't think it's perfect but I do think it's a workable first attempt, and it was fun to try it out and to learn some things about character training in Fizgig.
So far I've found it easiest to preserve identity with base Krea2 - you can add NSFW or refusal reduction / bypass LoRAs at partial strength and get a good effect - or by using the CielBleu checkpoint or the RawGirl v2 checkpoint. It feels weird that some of my other favorite checkpoints mess with her face so much, but that is an advanced character LoRA topic for me to learn more about some other day.
In the meantime you can reduce ID drift by mentioning her identity markers in the prompt (black hair, very dark hair, blue eyes, etc) or by adding the optional trigger word 'arskiara' (the trigger effect is quite small though tbh)
It's also weird that I'm posting Kiara without any clear idea whether she will be of use to anyone else; very different from my other LoRAs! I look forward to finding out, and input is welcome as always.
Happy generating!
Description
First try! The trigger is 100% optional.
FAQ
Comments (6)
Super cute! Has some Margo Robie / Emma Mackey qualities about her face (I like it!).
I found that facial drift comes from undertraining on the face (data set should be 75%+ with face - not all up close, but still in the shot). My Characters that have lower drift are usually 1-2 epochs past what I consider a well trained model in a base checkpoint.
That being said, many of the NSFW and female focused merges HEAVILY skew faces and can even topple the stability of an over trained model (they carry way too much face data - specifically soft 18 year old faces). The models I found that work well to keep identity are the: Krea 2 Turbo, Intorealism Flashv1, Kromafine Turbo, Serendipity, and Solstice. Pair that with TextFusion Refusal and/or Better Boobs 3 to get some added NSFW without much or any compromise to face.
Also noticing the file size, you MAY have a low enough rank that it gets stomped on if it's got any under trained nature to it.
@xFennec777 thanks man! Appreciate the expert input. The other thing I experimented with was the speed-optimized low-rank path in Fizgig with the per-image LR easing based on face ID similarity. Fun stuff to play with but I am a monkey pushing levers compared to making a style. I had an 80-image dataset of which half were face closeups. I will try your suggested checkpoints - thanks!
@arsibalt Not a problem! I have an article published that outlines a minimal / sweet-spot set of images. It requires as much consistency of character as possible across the set (my Faye character had some proportion and face shifts that caused noticeable drift in Sick Ollie and others - which both showed the overtuned nature of the checkpoint and the faults in my LoRA). Faye was a NIGHTMARE to get right, I ended up meticulously rebuilding all of her training data, after messing about by creating multiple scripts that attempted to stablise the model.
It may be as simple as adjusting your data set and pruning/regenerating a couple images that have outliers on them. The article is based on my work for Carys (she is a REAL SOLID model that still gets mangled by a few checkpoints).
@xFennec777 I did read your article! I should go back and re-read it though, it's dense with info
@arsibalt Just look for the section on dataset outlining (X number of this type) - should get you going!