Inter-Edit-Test

Official test benchmark release for the CVPR 2026 paper:

Inter-Edit: First Benchmark for Interactive Instruction-Based Image Editing

This repository hosts the public release of Inter-Edit-Test, a human-annotated benchmark for the Interactive Instruction-based Image Editing (I^3E) task.

Each sample contains:

• a source image,
• a coarse user-style interaction mask,
• a concise editing instruction,
• and a ground-truth edited image.

To simplify large-scale distribution on Hugging Face, image assets are packaged as tar archives while metadata remains directly accessible.

Highlights

• 6,250 human-annotated test pairs
• Bilingual instructions: English and Chinese
• User-style coarse masks, rather than segmentation-perfect masks
• High-resolution benchmark
• Challenging subsets including artistic styles, low-resolution images, low-aesthetic images, and ambiguous multi-instance edits
• Sanitized public release: all filenames are re-indexed to remove source-specific naming information

Release Statistics

Language distribution

• English: 3,413
• Chinese: 2,837

Edit type distribution

• Remove: 1,641
• Add: 1,603
• Local: 1,533
• Text Editing: 1,084
• Texture: 389

Repository Structure

Inter-Edit-Test-HF/
├── source_images.tar
├── target_images.tar
├── masks.tar
├── metadata.json
├── metadata.jsonl
└── README.md

Archive contents use index-based sanitized names:

• source_images/00000_source.xxx
• target_images/00000_gt.xxx
• masks/00000_mask.xxx

Metadata Format

Each record in metadata.json / metadata.jsonl contains:

• index: zero-based dataset index
• source_path: relative path inside the extracted archive
• gt_path: relative path inside the extracted archive
• mask_path: relative path inside the extracted archive
• edit_type: edit category
• language: instruction language
• instruction: concise editing instruction

Quick Start

Extract the archives after download:

tar -xf source_images.tar
tar -xf target_images.tar
tar -xf masks.tar

Then load metadata as usual:

import json
from pathlib import Path

root = Path(".")
with open(root / "metadata.json", "r", encoding="utf-8") as f:
    meta = json.load(f)

sample = meta[0]
source = root / sample["source_path"]
target = root / sample["gt_path"]
mask = root / sample["mask_path"]
instruction = sample["instruction"]

Benchmark Intent

Inter-Edit-Test evaluates whether a model can:

1. correctly understand a concise edit instruction,
2. localize the intended region from an imprecise user mask,
3. perform the desired edit inside the target area,
4. preserve and harmonize the surrounding image naturally.

Project Page

• GitHub: https://github.com/Delong-liu-bupt/Inter-Edit

Citation

If you use Inter-Edit-Test in your research, please cite:

@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 (CVPR)},
  year      = {2026}
}