ZeyuJiang1/OrionEditBench

---
language:
- en
license: apache-2.0
size_categories:
- 10K<n<100K
task_categories:
- image-text-to-image
tags:
- image-editing
- multimodal
- diffusion
configs:
- config_name: image-text-to-image
data_files: metadata/metadata.json
---

# 🧩 OrionEditBench

**OrionEditBench** is a large-scale dataset for **cross-image editing**, where each sample is structured as:

> **(reference image(s), source image) → synthesis image**

It is designed to support multi-image conditioned generation and editing, allowing models to integrate visual information across inputs.
The dataset contains approximately **50K high-quality samples**, covering key editing scenarios including **attribute transfer**, **style alignment**, and **multi-image fusion**.
We first release the core subset focusing on **attribute transfer**, including **subject replacement** and **appearance transfer**, with both single- and multi-subject settings. Additional data is under active curation and will be released in future versions.

---

## News

- **2026.4.01**: The dataset has been released.
- **2026.4.02**: We provide the metadata for the sub-tasks of subject replacement and appearance transfer within attribute transfer, focusing on the AI-synthesized portion.

## Overview

| Data Type | Number of Samples |
| :---------------------- | :----------------- |
| Attribute Transfer | 10,664 |
| Fusion | - |
| Style Alignment | -|
| **Total** | **-** |

OrionEditBench focuses on **generalized cross-image editing tasks**, where models are required to:
- transfer attributes from reference images
- preserve structural or semantic content from source images
- generate coherent target outputs

The dataset consists of a mixture of:

- automatically constructed samples generated using advanced multimodal models
- curated samples derived from publicly available data

---

## Format Illustration
To reduce training complexity and memory overhead in multi-reference-based editing, where separate branches are typically required for reference, source, and noise initialization, we pre-compose multiple reference images into a single input so that they share a unified processing branch. For clarity, we provide two representative data formats: (1) foreground-isolated references with background removed, and (2) concatenated pairs of reference images resized to match the source dimensions.

<p align="center">
<img src="assets/example_foreground.png" width="40%">
<img src="assets/example_concat.png" width="40%">
</p>

<p align="center">
<em>Top: foreground-isolated format. Botton: concatenated source–target format.</em>
</p>

## Data Structure

Each sample in the dataset is stored in JSON format:

```json
{
"edit_prompt": "...",
"t2i_prompt": "...",
"source_image": "images/source/xxx.png",
"reference_image": "images/reference/xxx.png",
"output_image": "images/synthesis/xxx.png",
"width": 1344,
"height": 768,
}
```

To extract the dataset from split .tar archives:

```json
cd /path/to/reference-source-synthesis
cat source_part_*.tar | tar -xf -
```

After extraction, you may partition the dataset into training and testing splits based on your experimental needs.

## Resources

* **GitHub**: [cityuhkai/OrionEdit](https://github.com/cityuhkai/OrionEdit)
* **Model**: [OrionEdit-qwen](https://huggingface.co/ZeyuJiang1/OrionEdit-qwen)
* **Paper**: [OrionEdit: Bridging Reference and Source Images for Generalized Cross-Image Editing](https://github.com/cityuhkai/OrionEdit)

## Citation
If you find our dataset helpful, please consider citing our work:

```
@article{
> Our paper has been accepted to CVPR 2026.
The official citation will be released upon publication.
}
```