# Minionsmonsters2026
A style LoRA for Krea 2. Applies a consistent visual treatment without locking the subject matter.
## Usage
Include the trigger word *mm** in your prompt.
Recommended weight: 0.8. Start there and adjust — styles often read well below 1.0, and lower weights blend more naturally with other LoRAs.
## Promt Examples
Here are some prompt examples to generate minions:
"mm, A group of small, yellow, one-eyed creatures wearing goggles and overalls stand in a dim, industrial-looking space, their expressions ranging from curious to concerned, captured in a tight, eye-level shot with soft, diffused lighting."
"mm, A group of yellow, one-eyed creatures in goggles and overalls huddle closely, one holding a worn book, under dim, moody lighting that casts shadows across their faces, framed tightly to emphasize their expressions and the intimate, slightly chaotic scene."
"mm, A close-up of a yellow, one-eyed character with a large goggles-like eye, wearing denim overalls, set against a dimly lit, rustic interior with warm candlelight. The camera focuses tightly on the character’s expressive face, capturing a moment of wide-eyed surprise or excitement."
"mm, A group of small, yellow, one-eyed creatures in blue overalls stand in a row, with one in the foreground holding a metal pipe, set against a brick wall backdrop under warm, focused lighting, captured in a medium shot that emphasizes their synchronized, expressive faces."
"mm, A single yellow, one-eyed character with a stone-like eye and leather straps stands in the foreground, surrounded by identical figures in a row, all wearing white bandages and seated on wooden benches under an overcast sky. The camera frames the central figure tightly, emphasizing their expression and detailed costume against a muted, atmospheric background."
## Training
| | |
|---|---|
| Base model | Krea 2 (trained on Raw) |
| Type | Style LoRA |
| Dataset | 120 images |
| Steps | 2000 |
| Network rank / alpha | 32 / 32 |
| Learning rate | 0.0002 |
| Optimizer | adamw8bit |
| Resolutions | 768 / 1024 |
## Generation settings
Trained on Krea 2 Raw, and transfers to Krea 2 Turbo for faster
generation.
- Raw: ~30 steps, guidance 3.5
- Turbo: ~8 steps, guidance 1.0
Description
FAQ
Comments (2)
i love it, are you trainning on entire movies? because i did a dataset of the minecraft movie, with 331 hand picked shots, my results still bad after 7 hours of training.
ohh wow. 7h training is to much. In the training process I skip the sampler part, if you do so, you can save around 30% of training time. Im my experience (I can be wrong) is no need to have 300+ images. between 60-120 is fine enought. One of the problems that you can have when training krea 2 is captioning. If the camptions in the data set are not good enought, you will always have a bad results. I use Qwen3-VL-8B-Instruct to caption my images if you use the 4B version is not good enought. If you are doing it manually, probably you are missing some important parts of the image to describe. One more thing to have in mind, if your traning data set, doesnt have for example a car, or a tree or a dog, and then you use your lora and you only prompt "triggerword, car" or "triggerword, tree" or "triggerword, dog" you are not going to see the image in the lora style. To achieve that you have to incresse the lora weight. Maybe 1.5, 2.0 or more.