VDEBench: Visual Document Editing Benchmark

Edit the requested text or table. Preserve the rest of the document.

VDEBench evaluates instruction-guided image editing on dense Chinese and English document pages. Models must make precise local changes while preserving surrounding text, table structure, typography, and page layout. The benchmark contains 3,506 examples, covering text replacement, addition, deletion, and table content and structure editing.

Read the paper · Get the code · Explore the data

Overview illustration from the VDE-Bench GitHub repository.

At a glance

Examples	Text edits	Table edits	Languages	Task types
3,506	1,578	1,928	English · Chinese	7

Each example pairs an input image, an editing instruction, and a reference target image with editing-region annotations and source provenance. Available masks describe the allowed editing region and the pixels that actually changed.

Editing examples

The following input–target pairs come from this dataset. They show the editing tasks and reference targets, rather than predictions from a model under evaluation. Click an image to inspect it at full resolution.

Text replacement

Instruction: 将标题“国资背景基金情况”替换为“国资背景基金分析”
Replace the title “国资背景基金情况” with “国资背景基金分析”.

Input document	Reference target

Example 48234 · text configuration · text replacement.

Table content editing

Task: Replace several table headers, cell entries, and the conclusion according to one instruction.

Input document	Reference target

Example 46007 · table configuration · text modification.

Show the full editing instruction

修改表格内的文字内容：将表头第二列的“描述听到的声音的强弱变化”改为“观察到的现象”；将第一列的“实验”改为“测试”；将“轻轻拨动钢尺”改为“缓慢移动滑块”，“振动幅度小 声音弱（小）”改为“移动距离短 速度慢（低）”；将“用力拨动钢尺”改为“快速移动滑块”，“振动幅度大 声音强（大）”改为“移动距离长 速度快（高）”；将“我的发现”改为“结论”；并将最后一行关于音量的描述替换为“速度是由物体移动的距离和时间决定的，移动距离越长，速度就越快；移动距离越短，速度就越慢。”

Quick start

pip install -U datasets pillow

Load the benchmark with Hugging Face Datasets. Streaming lets you inspect examples without first downloading the entire collection.

from datasets import load_dataset

ds = load_dataset(
    "hongzhuyi/VDEBench",
    name="default",
    split="test",
    streaming=True,
)

example = next(iter(ds))
print(example["id"], example["instruction"])
example["image_input"].save("input.png")
example["image_output"].save("reference.png")

Choose name="text" for text edits or name="table" for table edits. Remove streaming=True to download and load the complete configuration. The repository is public; authentication is optional.

Images and available masks are embedded in Parquet, so this interface needs no custom loading script or archive extraction.

Tasks and coverage

Region	Editing task	Examples
Text	Replacement	575
Text	Addition	504
Text	Deletion	499
Table	Text modification	524
Table	Text addition	472
Table	Text deletion	472
Table	Structure editing	460
Total		3,506

All configurations expose the frozen benchmark through a single test split. The default configuration contains all 3,506 examples; text and table are its two subsets, not additional examples. When constructing new partitions, group examples by source document to prevent source overlap.

Evaluation code

The VDE-Bench repository provides data-generation and evaluation tools. The evaluation protocol combines local edit-region assessment with global full-page assessment to examine both the requested change and its surrounding document context.

• OCR and structure evaluation: block matching with IoU, text metrics including CDM and BLEU, and TEDS for tables.
• Image-quality evaluation: PSNR, SSIM, LPIPS, FID, and CLIP Score.
• Layout analysis: document-element extraction and layout statistics.

See the repository's installation and usage instructions for environment setup, OCR backends, and evaluation commands.

Data format

Fields	Contents
id, instruction	Persistent example ID and editing instruction
region, edit_type, language, data_source	Editing task, language, and document category
image_input, image_output	Input and reference images, decoded as PIL images
allowed_mask, actual_difference_mask	Optional editing-region and changed-pixel masks; missing masks are null
label_input_json, label_output_json	JSON-encoded annotation structures
record_json, provenance_json, annotation_json	Compatibility records and provenance sidecars
source_dataset, source_group	Source identity and grouping
image_input_path, image_output_path	Relative image paths in the compatibility archives

Annotation rectangles use percentages of the recorded annotation canvas, which can differ from the image's native dimensions. The allowed editing region and actual changed pixels have separate annotations. See DATA_DICTIONARY.md for field definitions and coordinate conventions.

Original JSON/image interface

The root-level text_info_all.json and table_info_all.json retain the VDE-compatible records. Input, target, and mask files are stored in compatibility/assets-*.tar.

from pathlib import Path
import tarfile
from huggingface_hub import snapshot_download

root = Path(snapshot_download("hongzhuyi/VDEBench", repo_type="dataset"))
for archive in sorted((root / "compatibility").glob("assets-*.tar")):
    with tarfile.open(archive) as tar:
        tar.extractall(root, filter="data")

Use a Python version supporting tarfile extraction filters (Python 3.12+ recommended). load_dataset.py supports local JSON filtering. asset_manifest.json describes unpacked assets; files_sha256.json records repository file hashes except its own.

Sources and licensing

Source datasets include OmniDocBench, DocGenome, and DocLayNet, alongside the recorded source papers.

The data follows source-specific terms. See SOURCE_LICENSES.md and provenance_and_validation.json for attribution and reuse conditions.

Citation

If you use VDEBench in your research, please cite the paper:

@article{yi2026vdebench,
  title   = {VDE Bench: Evaluating The Capability of Image Editing Models to Modify Visual Documents},
  author  = {Yi, Hongzhu and Yang, Yujia and Wang, Yuanxiang and Li, Tong and Guan, Zhenyu and Zong, Tianyu and Chen, Jiahuan and Bao, Chenxi and Yang, Tiankun and Jin, Haopeng and Yuan, Yixuan and Wang, Xinming and Yu, Tao and Gao, Ruilin and Tao, Ruiwen and Liang, Haijin and Ma, Jin and Luo, Jinwen and Yeshani and Zuo, Xinyu and Xu, Jungang},
  journal = {arXiv preprint arXiv:2602.00122},
  year    = {2026},
  url     = {https://arxiv.org/abs/2602.00122}
}