Sign In

Not recently loaded

Download

1 variant available

bf16 SafeTensor

arcane_krea2.safetensors

BF16, good balance • 109.03 MB

Verified:

Type
LoRA
Stats

26

0

Reviews
Published

Oct 21, 2026

Base Model

Krea 2

Training
Steps: 2,000
Epochs: 4
Usage Tips
Strength: 1
Hash
AutoV2
9132CBFA26
Trigger Words
arcane (arknights), arknights: endfield, 1girl, solo, forehead mark, facial mark, grey hair, multicolored hair, long hair, blue eyes, black gloves, cape
grid_all_checkpoints.jpg
# Arcane (Arknights: Endfield) – Krea 2 LoRA

Character LoRA for **Arcane** from *Arknights: Endfield*, trained on Krea 2 Raw. Works on Krea 2 Raw, Krea 2 Turbo and Turbo merges such as Dasiwa.

## Trigger words

```
arcane (arknights)
arknights: endfield
```

## Recommended prompt

```
arcane (arknights), arknights: endfield, 1girl, solo, forehead mark, facial mark, grey hair, multicolored hair, long hair, blue eyes, black gloves, cape
```

Optional outfit tags: `elbow gloves, two-tone gloves, leotard, red thighhighs, black cape, ponytail`

## Usage

| | Krea 2 Turbo (and merges) | Krea 2 Raw |
|---|---|---|
| LoRA strength | 1.0 | 1.0 |
| Steps | 8 | 20–30 |
| CFG | 1 | about 4 |
| Sampler / scheduler | euler / simple | euler / simple |

- Eye colour varies, so name it in the prompt if you want a specific one.
- The LoRA only activates with the trigger word; prompts without it are unaffected.

## Training

- Base model: `krea/Krea-2-Raw`
- Dataset: 60 images (Danbooru-style tags) plus a regularization set
- 2000 steps, AdamW, lr 1e-4 constant, weight decay 1e-4
- Batch 1 × gradient accumulation 2, 512/768 px
- Trained on an Intel Arc B580 (12 GB) with [ai-toolkit-xpu](https://github.com/SHoogstad/ai-toolkit-xpu) v1.2.0

## Civitai tags

```
anime, character, arknights, arknights endfield, arcane, girl, video game, krea 2, krea, lora, woman, game character
```