mohan2/StyleTransfer-Reward-StyleScore

---
license: apache-2.0
task_categories:
- image-to-image
tags:
- style-transfer
- reward-model
- preference-learning
size_categories:
- 100K<n<1M
---

# StyleTransfer-Reward-StyleScore

Complete Reward Model training data from **StyleScore evaluation pipeline**.

## 📊 Data Contents (~292GB total)

### Core Reward Data
| Archive | Size | Description |
|---------|------|-------------|
| `style_images.tar.part_*` | 123GB | Winner images (高质量风格化结果) |
| `genref_wds_content.tar.part_*` | 122GB | Content images (原始内容图) |
| `loser_images.tar` | 14GB | Loser images (低质量对比) |

### Pair Configuration
| Archive | Size | Description |
|---------|------|-------------|
| `cnt_sty_pairs_cfg.tar` | 3.4GB | 100k pairs 配对配置 (JSONL) |
| `omnistyle_content.tar` | 2.6GB | OmniStyle content 图片 |

### Style References
| Archive | Size | Description |
|---------|------|-------------|
| `style30k_processed.tar` | 3.4GB | Style30K 风格参考图库 |
| `wikiart_general_processed.tar` | 4.9GB | WikiArt 评测数据 |
| `style_image_synthesis.tar.part_*` | 19GB | 风格合成数据 |

## 🚀 Usage

```bash
# Merge split archives
cat style_images.tar.part_* > style_images.tar
cat genref_wds_content.tar.part_* > genref_wds_content.tar

# Extract all
for f in *.tar; do tar -xvf "$f"; done
```

## 📝 Pair Format

Training pairs in `cnt_sty_pairs_cfg/pairs_run_100k.jsonl`:
- `winner`: Higher StyleScore (better style transfer quality)
- `loser`: Lower StyleScore (worse quality)
- `content`: Original input image