Updated: May 3, 2026
style
# Mretsis Style LoRA — 半厚涂画风 LoRA
> 🇨🇳 中文 ↓ | 🇬🇧 English ↓↓
---
## 中文介绍
这是我训练的第一个 LoRA，还有很多不足之处，敬请谅解。
### 初衷
我的本意是制作一个用于改善工作流画风的 Style LoRA——追求一种**有轻微厚涂质感、但没有油腻 AI 感**的渲染风格，同时让画面形成**人物与背景主次分明、详略得当**的层次关系。简而言之：清爽、有笔触、不塑料。
### 版本说明
v1 — 风格-角色纠缠版
v1 在训练过程中出现了「风格-角色纠缠」(style-character entanglement) 的问题：由于训练数据集中角色多样性不足（原作 CG 女主基本只有一个人），模型将画风和角色外观绑定在了一起，导致在非训练角色上泛化能力很差——换一个角色，画风就跑偏了。
不过，如果你恰好需要的话，也可以把 v1 当作一个 Mretsis 游戏女主角色 LoRA 来使用。输入相应的角色提示词（白/灰发、蓝瞳、军装等），应该能较好地还原。
v2 — 泛化改进版
v2 通过以下策略显著改善了泛化能力：
- 数据集精选（删除冗余构图，保留姿势/光线/场景多样性）
- Caption 增强（随机替换角色外观描述，打破视觉-文本关联）
- 提高 Caption Dropout Rate（迫使模型学习视觉模式而非文本关联）
- 详尽的角色外观描述（让文本编码器承担角色信息，减轻 UNet 负担）
v2 在未训练角色上能保持一致的画风表现。
### 使用建议
- 推荐权重: 0.8–1.2（建议从 0.8 开始调整，游戏中的角色特征0.8，否则推荐1.2）
- 触发词: Mretsis
- 建议**配合其他画师串和 LoRA** 使用，形成一种画风 cocktail，而非单独依赖本 LoRA
- 如果在特定角色上画风偏弱，可以尝试：
- 提高权重到 0.9–1.2
---
## English Introduction
This is my first LoRA — there's still a lot of room for improvement, and I appreciate your understanding.
### Purpose
I built this as a style LoRA to refine my generation workflow. The goal: a rendering style with subtle impasto texture without the greasy AI look, combined with a clear visual hierarchy between subject and background — clean, brushy, not plastic.
### Version Notes
v1 — Style-Character Entanglement
v1 suffers from style-character entanglement: the training dataset lacked character diversity (the source material features essentially one protagonist), so the model fused the rendering style with the character's appearance. Generalization to non-training characters is poor — switch the character and the style drifts.
That said, v1 can still serve as a character LoRA for the Mretsis game protagonist. Use the appropriate character tags (white/grey hair, blue eyes, military attire) for best results.
v2 — Improved Generalization
v2 significantly improves generalization through:
- Dataset curation (removed redundant compositions, preserved pose/lighting/scene diversity)
- Caption augmentation (randomly swapped character appearance descriptors to break visual-text associations)
- Increased caption dropout rate (forces the model to learn visual patterns rather than text associations)
- Exhaustive character captioning (offloads character identity to the text encoder, reducing UNet entanglement)
v2 maintains consistent style across unseen characters.
### Usage Recommendations
- Recommended weight: 0.8–1.2 (start at 0.8, if not character in Mretsis's game, try 1.2)
- Trigger word: Mretsis
- Best used as part of a style cocktail with other artist tags and LoRAs, rather than standalone
- If style is weak on a particular character, try:
- Increasing weight to 0.9–1.2
---
## Technical Details
### Base Model
| Item | Value |
|------|-------|
| Checkpoint | chenkinNoobXLCKXL v02 (NoobAI-XL / Illustrious-XL lineage) |
| VAE | sdxl-vae-fp16-fix |
### Training Parameters
| Parameter | v1 | v2 |
|-----------|----|----|
| Dataset size | ~70 images | ~45 images (curated) |
| Network type | LoRA (Standard) | LoRA (Standard) |
| Network dim (rank) | 32 | 32 |
| Network alpha | 16 | 16 |
| Learning rate | 1e-4 | 1e-4 |
| UNet LR | 1e-4 | 1e-4 |
| Text Encoder LR | 5e-5 | 5e-5 |
| Optimizer | AdamW8bit | AdamW8bit |
| LR scheduler | cosine_with_restarts | cosine_with_restarts |
| Scheduler cycles | 1 | 3 |
| Warmup steps | 0 | ~10% of total steps |
| Epochs | 20 | 14 |
| Total steps | ~1,400 | ~2,400 |
| Batch size | 1 | 1 |
| Resolution | 1024 | 1024 |
| Buckets | ON (512–2048, step 64) | ON (512–2048, step 64) |
| Noise offset | 0.05 | 0.1 |
| Min SNR gamma | 5 | 5 |
| Caption dropout | 0 | 0.1 |
| Shuffle caption | ON | ON |
| Keep tokens | 1 | 1 |
| Clip skip | 2 | 2 |
| Mixed precision | fp16 | fp16 |
| Max token length | 225 | 225 |
### Key Differences v1 → v2
- Dataset curated from ~70 → ~45 images (removed redundant compositions)
- Caption augmentation: 30% random swap on character-specific tags (hair color, eye color, body type)
- Caption dropout rate: 0 → 0.1
- LR warmup: 0 → 10% of total steps
- Cosine restart cycles: 1 → 3
- Noise offset: 0.05 → 0.1
- Exhaustive character descriptions in captions to reduce style-character entanglement
### Training Tools
- Training framework: Kohya SS (sd-scripts)
- Hardware: RunPod cloud GPU
- Captioning: Manual review + WD14 tagger baseline
---
First LoRA, lessons learned. Thanks for trying it out. 🌲