Updated: May 3, 2026
character
# Deepseek Whalegirl LoRA - Civitai Upload Description
## 中文说明
这是一个 Deepseek 鲸鱼娘人物 SDXL LoRA。角色的核心身份特征由 LoRA 自身学习，包括眼罩、长发、鲸鱼元素、白花装饰、项圈吊坠和纤细体型等；推荐 prompt 里不需要反复写这些固定特征。
### 触发词
```text
Deepseek whalegirl
```
### 推荐权重
默认推荐：
```text
0.80 - 1.20
```
如果眼罩、鲸鱼元素或发饰不稳定，可以提高到 0.85 - 0.90。如果裙装、背景或姿势太固定，建议降到 0.60 - 0.75。
### 仅触发词测试
```text
Deepseek whalegirl
```
### 推荐额外 Prompt
白背景全身测试：
```text
Deepseek whalegirl, full body, standing, white background, layered dress, white dress, blue dress, frilled dress, lace trim, blue ribbon, barefoot
```
半身表情测试：
```text
Deepseek whalegirl, upper body, looking at viewer, hands clasped, hands on chest, blush, parted lips, white background
```
海边场景测试：
```text
Deepseek whalegirl, full body, standing, ocean, beach, sunset, water, waves, reflection, barefoot, parted lips
```
坐姿测试：
```text
Deepseek whalegirl, full body, sitting, hugging knees, knees up, layered dress, frilled dress, lace trim, barefoot, blush
```
水下漂浮测试：
```text
Deepseek whalegirl, full body, floating, underwater, ocean, bubbles, jellyfish, arms out, one knee bent, barefoot
```
### 使用建议
只写 Deepseek whalegirl 应该能召唤角色本体。推荐额外 prompt 主要用于控制构图、背景、姿势、裙装状态和表情。
不建议在常规使用中额外写入核心身份特征，例如发色、眼罩、鲸鱼尾发饰、白花发饰、项圈、吊坠或体型词。那些特征已经交给 LoRA 学习；反复写入可能让角色变得更僵硬，也会降低换装、换背景和换姿势的灵活性。
## English Description
This is an SDXL character LoRA for a Deepseek whale girl. The character identity is learned visually by the LoRA, including the blindfold, long hair, whale motif, white flower ornament, choker pendant, and slender body shape. These fixed identity traits do not need to be repeated in the prompt.
The LoRA works well for ocean-themed, whale-themed, dreamy girl character generation. It can be used with white backgrounds, ocean scenes, night sky, moonlight, water reflections, indoor display spaces, and pose controls such as full body, upper body, sitting, kneeling, hugging knees, or floating.
### Trigger Word
```text
Deepseek whalegirl
```
### Recommended Weight
Default starting point:
```text
0.80 - 1.20
```
### Trigger-Only Test
```text
Deepseek whalegirl
```
### Recommended Extra Prompts
White background full-body test:
```text
Deepseek whalegirl, full body, standing, white background, layered dress, white dress, blue dress, frilled dress, lace trim, blue ribbon, barefoot
```
Upper-body expression test:
```text
Deepseek whalegirl, upper body, looking at viewer, hands clasped, hands on chest, blush, parted lips, white background
```
Ocean scene test:
```text
Deepseek whalegirl, full body, standing, ocean, beach, sunset, water, waves, reflection, barefoot, parted lips
```
Sitting pose test:
```text
Deepseek whalegirl, full body, sitting, hugging knees, knees up, layered dress, frilled dress, lace trim, barefoot, blush
```
Underwater floating test:
```text
Deepseek whalegirl, full body, floating, underwater, ocean, bubbles, jellyfish, arms out, one knee bent, barefoot
```
## Training Parameters
- Base model: chenkinNoobXLCKXL_v03.safetensors
- Architecture: SDXL LoRA
- Network module: networks.lora
- Dataset: 22 images
- Dataset folder: 4_Deepseek whalegirl
- Repeats: 4
- Total train images per epoch: 88
- Epochs: 20
- Total steps: 1760
- Batch size: 1
- Resolution: 1024x1024
- Bucketing: enabled
- Bucket range: 512 - 2048
- Bucket no upscale: enabled
- Network rank/dim: 32
- Network alpha: 16
- Network dropout: 0.03
- Learning rate: 1e-4
- U-Net LR: 1e-4
- Text encoder LR: 1e-5
- Scheduler: cosine_with_restarts
- Scheduler cycles: 3
- Warmup steps: 100
- Optimizer: AdamW8bit
- Mixed precision: fp16
- Clip skip: 2
- Caption shuffle: enabled
- Keep tokens: 1
- Max token length: 225
- Caption dropout rate: 0
- Caption tag dropout rate: 0.05
- Min SNR gamma: 5
- Noise offset: 0.03
- Cache latents: enabled
- Cache latents to disk: enabled
- Gradient checkpointing: enabled
- Flip augmentation: disabled