💃 Momo — Gyaru-Style Character LoRA for Krea 2
A character LoRA for Krea 2 — a fictional Japanese gyaru-style woman with blonde hair and tanned/dark skin. Consistent look across poses and outfits, trained for 800 steps on 37 images.
⚠️ Requires Krea 2 as the base model. Won't work correctly on other checkpoints.
What's included:
🎨 Momo LoRA — the trained character model,
.safetensors
Trigger word: Japanese young woman, blonde hair, dark skin
Recommended settings:
Strength:
[0.7–1.0](adjust to taste)Sampler:
12 steps in Turbo modelsWorks well with:
most lighting and ambient loras
Quick start:
Drop the
.safetensorsfile into yourlorasfolderLoad it with a standard LoRA loader node hooked to Krea 2
Include
Japanese young woman, blonde hair, dark skinin your prompt
Heads up: First locally trained Lora I've ever done (previous ones were made in websites), so, the Lora is not perfect, but I think it's quite usable.
Description
FAQ
Comments (11)
Very nice first model. Thank you & welcome to Krea 2!
Thanks, more models to come soon!
you should totally make all the kawaii girl style loras
Thanks, I still need to fix some mistakes in my Lora making, once I do that, I'll come back with more characters.
Honestly this LORA character kept rendering completely differently most of the time, and looked - well let's say - not very similar to your screenshots.
Was fun to make a few Gyaru galleries, though ;)
Sorry to hear, try pushing the lora to 1.4 strength, that works for me when some character drift appears.
I think the problem wasn't that it renders too weak / generic, but that her face was way too round, like a puffed-up pancake - while the used K2 checkpoint renders asians slenderly in general. And it kept showing faces from slightly different people, until i started prompting more specifically.
There's just one image where her face looked much closer to your images, that's the one where she stands with that XXXL name-t-shirt and pouts https://civitai.red/images/142415438
But yeah thanks for sharing, the art inspiration was pretty good.
@ArttTaku BTW - do you render with Turbo checkpoints? (the images don't state it)
Because Turbo has a rather drastic normalizing effect. Try a checkpoint like Sick Ollie https://civitai.red/models/2676616/sick-ollie?modelVersionId=3154408 (i use the INT8 version with the FP8 text encoder). This also listens very closely to prompts for skintone, etc.
@LurkingAI Yeah, I only use Turbo models for my renders... but i figured this Lora would work with all Krea2 models?
@ArttTaku I think Turbo reacts very differently to LORAs and prompts.
I used a lot of DMD / Turbo for SDXL and it was lightning-fast, but the LORA models either worked in Turbo or Normal SDXL - but usually not both. So you might get a crisp, proper replication for a given LORA in Turbo, but it's a mess in regular. And vice versa.
In Krea2 Turbo, the memory demands are ironically larger (as even the INT8 version needs the full-size text encoder which is twice as big) so i stick with regular INT8 checkpoint.
Well, it feels like Turbo just grabs the fastest associations in their datasets, so you always get a result which looks like existing photos. But it doesn't actually listen entirely to what the prompts or the LORAs are defining. While regular Krea2 goes a really long way trying to render exactly what you prompted.
I think this is also why all the Krea Turbo images look relatively good and authentic, but they have this normalized look where everything gravitates towards the ideal 25-year-old photomodel with more idealized features. It also makes the images look more desaturated and amateurish, though that seems like part of the deal to make the Turbo images look "real".
(You can still get amateur styles with regular Krea2 though i despise this low-grade smartphone / webcam style of the typical Turbo render).
@LurkingAI Thanks for all the info, I'll keep it in mind for my next loras.