# Krea 2 Natural Expressions
I made this LoRA to fix the stiff, robotic face look and generic centering that often happens on Krea 2.
It helps generate subtle, organic micro-emotions (*genuine laughs, hesitation, smirks, thoughtful glances, tired looks*) and creates more dynamic, candid framing without forcing a specific art style.
## NEW: Version 2.0 is Live!
After the huge success and feedback on v1, **v2.0** brings a major overhaul:
*Total Prompt Fidelity**: Greatly reduces hallucinations, unwanted background objects, and excessive lighting bleed. The LoRA listens to your prompt with high precision.
*Hand-Tuned Dataset**: Dataset expanded with curated images and completely hand-refined captions to isolate facial emotions without locking the scene.
*Multi-Style Mastery**: Exceptional on photorealism, but now **stronger than ever on anime, 3D, digital art, and stylized illustrations**.
## Two Editions Available in v2.0:
When downloading v2.0, you have two .safetensors files available:
1. **Standard Edition krea2_natural_expressions_V2.safetensors)** (Default / Primary):
* Focuses purely on natural facial expressions, eye contact, and emotional subtleties while leaving your original lighting, clothing, and background strictly untouched.
2. **Cinematic Atmosphere Edition ..._Cinematic_Atmosphere.safetensors)** (Additional File):
* For creators who want a bolder, high-impact visual look. In addition to facial expressions, it deepens contrast and re-grades the overall scene lighting. It carries forward the punchy, dramatic spirit of v1 with v2's enhanced stability.
## How to Use:
* **Trigger Word**: **None!** Just describe the expression and pose naturally in your prompt (*"subtle smirk", "candid laughing portrait", "intense thoughtful look"*).
* **Recommended Weight**: 0.6 to 1.2 (Sweet spot is between *0.8 and 1.0**).
* **Base Model**: Krea 2 (Raw / Turbo).
## Community & Feedback:
Please feel free to share your generations and prompts in the gallery below! I love seeing what the community creates and your feedback directly helps me improve future updates. Enjoy ! ❤️
Description
Initial release! Feel free to experiment with different weights (0.6 to 1.0). It's designed to bring more dynamic framing, natural micro-emotions, and candids without rigid poses.
FAQ
Comments (6)
It has a habit of adding tattoos to the characters in my images.
Okay, that's good to know. I haven't trained the model with many tattooed people, I'll look into that for version 2! Thanks for your comment !
Use Krea2 Prompt Weight (KJNodes) instead of the default CLIP Text Encode.
This way you can prompt (tattoos:-1).
It conveys emotions very well, but it constantly adds all sorts of details, effects, lighting, and things you didn't ask for. It especially likes to add fire to faces—literally flames on faces—especially on sickOllie. Reducing the weight to 0.6 doesn't really help. The problem occurs at every checkpoint: sickOllie, museYogi3.5, or moody7.0. When using the museYogi3.5 model, there’s constantly semen on the face or steam coming out of the mouth. This LORA is just some kind of uncontrollable chaos. =(
@killbrain Haha thanks for the honest feedback! On one point you're totally right: the LoRA can indeed slightly bleed into overall scene lighting and atmosphere—I'm actually working on refining that right now for V2 to isolate expressions much better.
For the rest though, rest assured there is zero fire or explicit fluids in my training data! I trained this LoRA on theater actors and royalty-free art/drawing reference models. There are no explicit scenes in my dataset—sometimes artistic nudes, but nothing more. 😂
What you're experiencing really comes from the base checkpoints you're running: museYogi is NSFW-biased, while sickOllie is probably similar. Obviously, I couldn't test the LoRA across every single community fine-tune out there, but it's always super valuable to get this kind of feedback to find areas for improvement.
To tame the chaos: try it on a cleaner base (like vanilla Krea 2 or OurSecret) and keep your CFG low around 1.0, as Krea 2 blows up effects very easily with higher CFG.
Let me know if that helps!
@Guillaume127 My guess is that it's related to captioning - to make sure you're training the expression only and not the look of the person - you have to describe things in the caption that you DON'T want the trainer to put in the lora. So if I were to try and do this I would use one person with many expressions to that the learning is focusing on the variations between them. You have to negate everything you don't want to be trained in the captions, so mentioning the hair colour in each image means the hair colour is not the thing that the trainer is "focusing on". Sorry if this is already obvious but I think that captioning is not something to be left up to a captioning model, they're terrible.