a1557811266/Inter-Edit-Train

---
language:
- en
- zh
pretty_name: Inter-Edit-Train
size_categories:
- 1M<n<10M
task_categories:
- image-to-image
tags:
- image-editing
- benchmark
- computer-vision
- multimodal
---

# Inter-Edit-Train

Inter-Edit-Train is the official large-scale training set released for the CVPR 2026 paper **Inter-Edit: First Benchmark for Interactive Instruction-Based Image Editing**.

This dataset is designed for the Interactive Instruction-based Image Editing (I^3E) task, where a model performs localized image edits from a concise textual instruction together with imprecise spatial guidance.

## Highlights

- **1,099,964** image editing pairs
- **610,186** unique source images
- Four edit types: **Local**, **Add**, **Remove**, and **Texture**
- Seven common aspect ratios from **16:9** to **9:16**
- Includes edit instructions, masks, bounding boxes, and an extra `better_data` flag
- All release filenames are anonymized with index-based names; original internal filenames are not exposed in the packaged archives

## Relation to the paper

This release corresponds to the training split described in the paper. The released manifest keeps the fields needed for training and data usage:

- `instruction`
- `edit_type`
- `bounding_box`
- `bbox_reference_dimensions`
- `better_data`
- anonymized source / target / mask asset locations

The key `better_data` is **not** a paper-defined benchmark field. It is an additional release-only flag indicating samples that were judged to be more suitable for training after filtering.

## Data layout

Because the full training set is extremely large, the assets are released as sharded tar archives:

- `source_shards/source-xxxxx-of-xxxxx.tar`
- `asset_shards/asset-xxxxx-of-xxxxx.tar`
- `metadata/train-xxxxx-of-xxxxx.jsonl.gz`

Each asset name inside the tar archives is anonymized:

- source image: `sources/source_0000000.png`
- edited image: `targets/target_0000000.png`
- mask image: `masks/mask_0000000.png`

Each metadata row records which tar shard and which internal filename should be used for that sample.

## Metadata schema

Each JSONL record contains:

- `sample_id`: zero-based sample index
- `source_id`: zero-based unique source-image index
- `edit_type`
- `instruction`
- `better_data`
- `bounding_box`
- `bbox_reference_dimensions`
- `source_archive`
- `source_file`
- `asset_archive`
- `target_file`
- `mask_file`

## Example metadata entry

```json
{
"sample_id": 0,
"source_id": 0,
"edit_type": "Add",
"instruction": "添加一双发光的筷子",
"better_data": false,
"bounding_box": [357, 694, 902, 926],
"bbox_reference_dimensions": {"width": 960, "height": 960},
"source_archive": "source_shards/source-00000-of-00245.tar",
"source_file": "sources/source_0000000.png",
"asset_archive": "asset_shards/asset-00000-of-00275.tar",
"target_file": "targets/target_0000000.png",
"mask_file": "masks/mask_0000000.png"
}
```

## Usage notes

- This is the **training** release, not the manually annotated test benchmark.
- The canonical sample order follows the original `Inter-Edit-train.json`.
- Source images are deduplicated globally and indexed separately from sample indices.
- Asset filenames are anonymized by design.

## Citation

If you use this dataset, please cite:

```bibtex
@inproceedings{liu2026interedit,
title={Inter-Edit: First Benchmark for Interactive Instruction-Based Image Editing},
author={Liu, Delong and Hou, Haotian and Hou, Zhaohui and Huang, Zhiyuan and Han, Shihao and Zhan, Mingjie and Zhao, Zhicheng and Su, Fei},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year={2026}
}
```