undefined443/JourneyDB-recaption

---
license: cc-by-nc-4.0
task_categories:
- image-to-text
- text-to-image
language:
- en
tags:
- journeydb
- midjourney
- recaption
- vision-language
- ai-generated
size_categories:
- 1M<n<10M
configs:
- config_name: default
data_files:
- split: train
path: data.parquet
---

# JourneyDB Recaption

Recaptioned version of the [JourneyDB](https://huggingface.co/datasets/JourneyDB/JourneyDB) dataset using Qwen vision-language models.

## Dataset Description

JourneyDB is a large-scale dataset of AI-generated images from Midjourney. This recaptioned version provides detailed visual descriptions generated by a vision-language model, which are more accurate than the original generation prompts for describing actual image content.

### Statistics

| Metric | Count |
| ---------- | ----------------- |
| Total rows | 3,389,605 |
| File size | ~246 MB (Parquet) |

### Columns

| Column | Type | Description |
| ----------------- | ------ | -------------------------------------- |
| `img_path` | string | Relative path to image in JourneyDB |
| `width` | int | Image width in pixels |
| `height` | int | Image height in pixels |
| `aesthetic_score` | float | Aesthetic score (may be null for some) |
| `caption` | string | Generated visual description |
| `model` | string | Model used for recaptioning |

### Recaption Models

| Model | Count |
| --------------------------------------------------------------------------------- | --------- |
| [Qwen/Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) | 2,249,747 |
| [Qwen/Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) | 1,139,858 |

- **Prompt style**: COCO-style short caption ("Describe this image in one simple sentence...")

### Preprocessing

Before recaptioning, the following filters were applied to the original JourneyDB dataset:

- **Resolution filter**: min(width, height) >= 512
- **Aspect ratio filter**: max(width, height) / min(width, height) <= 2.0

### Example

```python
{
"img_path": "./000/728deb7c-a5e2-463c-8f75-5f62dae521ac.jpg",
"width": 1024,
"height": 1024,
"aesthetic_score": 7.276855,
"caption": "A girl sits at a table with books, looking directly at the camera.",
"model": "Qwen/Qwen2.5-VL-7B-Instruct"
}
```

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("undefined443/JourneyDB-recaption")
```

Access the data:

```python
for sample in dataset["train"]:
img_path = sample["img_path"]
caption = sample["caption"]
width, height = sample["width"], sample["height"]
aesthetic_score = sample["aesthetic_score"]
# Use with JourneyDB images
```

## License

This dataset inherits the license from JourneyDB (CC-BY-NC-4.0).

## Related

- [JourneyDB](https://huggingface.co/datasets/JourneyDB/JourneyDB) - Original dataset
- [Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) - Vision-language model used for recaptioning
- [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) - Vision-language model used for recaptioning

license: CC-BY-NC-4.0（知识共享署名-非商业性使用4.0国际许可协议）
task_categories:
- 图像到文本（image-to-text）
- 文本到图像（text-to-image）
language:
- 英语（en）
tags:
- JourneyDB（JourneyDB）
- Midjourney（Midjourney）
- 重标注（recaption）
- 视觉语言（vision-language）
- AI生成（ai-generated）
size_categories:
- 100万<样本量<1000万
configs:
- config_name: default
data_files:
- split: train
path: data.parquet

# JourneyDB重标注数据集
本数据集是基于[JourneyDB（JourneyDB）](https://huggingface.co/datasets/JourneyDB/JourneyDB)原始数据集，使用Qwen视觉语言（vision-language）模型完成重标注的版本。

## 数据集说明
JourneyDB（JourneyDB）是一个源自Midjourney（Midjourney）的大规模AI生成图像数据集。本次重标注版本提供了由视觉语言模型生成的精细化视觉描述，相较于原始生成提示词，该描述更贴合图像的实际内容，准确性更优。

### 统计信息
| 指标 | 数值 |
| ------------ | ------------------- |
| 总样本数 | 3,389,605 |
| 文件大小 | 约246 MB（Parquet格式） |

### 字段说明
| 字段名 | 类型 | 描述说明 |
| ----------------- | ------ | --------------------------------------- |
| `img_path` | 字符串 | 图像相对于JourneyDB（JourneyDB）数据集根目录的相对路径 |
| `width` | 整数 | 图像宽度，单位为像素 |
| `height` | 整数 | 图像高度，单位为像素 |
| `aesthetic_score` | 浮点数 | 图像美学评分（部分样本可能为空） |
| `caption` | 字符串 | 模型生成的视觉描述文本 |
| `model` | 字符串 | 用于生成该条目标注的重标注模型名称 |

### 重标注模型
| 模型名称 | 样本数量 |
| --------------------------------------------------------------------------------- | ----------- |
| [Qwen/Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) | 2,249,747 |
| [Qwen/Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) | 1,139,858 |

- **提示词风格**：采用COCO（COCO）风格短描述，提示词格式为“用一句简洁的语句描述该图像……”

### 预处理流程
在进行重标注前，我们对原始JourneyDB（JourneyDB）数据集应用了以下筛选规则：
- **分辨率筛选**：图像的短边尺寸≥512像素
- **宽高比筛选**：图像长边与短边的比值≤2.0

### 示例数据
python
{
"img_path": "./000/728deb7c-a5e2-463c-8f75-5f62dae521ac.jpg",
"width": 1024,
"height": 1024,
"aesthetic_score": 7.276855,
"caption": "一名女孩坐在摆放着书籍的桌前，直视镜头。",
"model": "Qwen/Qwen2.5-VL-7B-Instruct"
}

## 使用方法
python
from datasets import load_dataset

dataset = load_dataset("undefined443/JourneyDB-recaption")

访问数据示例：
python
for sample in dataset["train"]:
img_path = sample["img_path"] # 获取图像相对路径
caption = sample["caption"] # 获取重标注描述文本
width, height = sample["width"], sample["height"] # 获取图像分辨率
aesthetic_score = sample["aesthetic_score"] # 获取图像美学评分
# 可结合原始JourneyDB（JourneyDB）数据集的图像进行使用

## 许可协议
本数据集沿用JourneyDB（JourneyDB）的许可协议：CC-BY-NC-4.0（知识共享署名-非商业性使用4.0国际许可协议）。

## 相关资源
- [JourneyDB（JourneyDB）](https://huggingface.co/datasets/JourneyDB/JourneyDB) - 原始JourneyDB（JourneyDB）数据集
- [Qwen/Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) - 本次重标注使用的视觉语言模型
- [Qwen/Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) - 本次重标注使用的视觉语言模型