---
license: apache-2.0
task_categories:
- image-to-image
language:
- en
size_categories:
- n<1K
---
arxiv: arxiv.org/abs/2509.04582

# DragBench Dataset
A benchmark dataset for drag-based image editing evaluation, consisting of two subsets with additional annotations for [Inpaint4Drag](https://visual-ai.github.io/inpaint4drag/) method evaluation.


## Dataset Overview

DragBench provides standardized evaluation data for interactive point-based image editing research. The dataset combines:
- **DragBench-DR** (205 samples): Modified version of DragBench-D from [DragDiffusion](https://github.com/Yujun-Shi/DragDiffusion/)
- **DragBench-SR** (100 samples): Modified version of DragBench-S from [SDE-Drag](https://github.com/ML-GSAI/SDE-Drag)

Both subsets include additional annotations for Inpaint4Drag method evaluation.

## Dataset Structure

```
drag_data/
├── dragbench-dr/          # DragBench-DR (205 samples)
│   ├── animals/
│   │   ├── sample_001/
│   │   │   ├── original_image.png
│   │   │   ├── meta_data.pkl
│   │   │   ├── user_drag.png
│   │   │   ├── meta_data_i4p.pkl
│   │   │   └── user_drag_i4p.png
│   │   └── ...
│   └── [other_categories]/
└── dragbench-sr/          # DragBench-SR (100 samples)
    ├── sample_001/
    │   ├── original_image.png
    │   ├── meta_data.pkl
    │   ├── user_drag.png
    │   ├── meta_data_i4p.pkl
    │   └── user_drag_i4p.png
    └── ...
```

## File Descriptions

Each sample contains the following files:

### Core Files
- **`original_image.png`**: The input image to be edited
- **`meta_data.pkl`**: Annotation data containing editing instructions and metadata
- **`user_drag.png`**: Visualization of the original drag editing annotation

### Inpaint4Drag Annotations
- **`meta_data_i4p.pkl`**: Additional annotation for Inpaint4Drag method evaluation
- **`user_drag_i4p.png`**: Visualization of the Inpaint4Drag (I4P) annotation

## Metadata Format

### `meta_data.pkl`
```python
{
    'prompt': str,  # Text prompt describing the desired output image
    'points': list, # Point coordinates [(x1,y1), (x2,y2), ..., (xn,yn)]
                   # Format: handle_point1, target_point1, handle_point2, target_point2, ...
    'mask': array  # Binary mask specifying the editable area
}
```

### `meta_data_i4p.pkl`
```python
{
    'points': list, # Point coordinates with same editing intent as original annotation
    'mask': array  # Binary mask specifying the deformable region for I4P method
}
```