📘 VisEditBench: A Benchmark for Visual Instruction-Based Diagram Editing

Overview

VisEditBench is a benchmark dataset for evaluating models on editing TikZ diagrams based on visual instructions.

Each sample consists of:

• a source diagram
• a target diagram
• a visual instruction
• TikZ code for both the source and target diagrams
• metadata describing the edit operation

The benchmark evaluates whether models can interpret visual instructions and make the corresponding changes to TikZ diagrams.

---

Dataset Structure

dataset/
├── metadata.jsonl
├── source_images/
├── target_images/
└── visual_instructions/

Files

• metadata.jsonl — Contains one JSON object per sample.
• source_images/ — Original diagrams (PNG).
• target_images/ — Edited diagrams (PNG).
• visual_instructions/ — Instruction images.

---

Metadata Format

Each line in metadata.jsonl has the following structure:

{
    "id": "00000069_STRUCTURE_1_edit_0",
    "source_image_file_name": "source_images/00000069_STRUCTURE_1_edit_0.png",
    "visual_instruction_file_name": "visual_instructions/00000069_STRUCTURE_1_edit_0.png",
    "target_image_file_name": "target_images/00000069_STRUCTURE_1_edit_0.png",
    "uri": "https://tex.stackexchange.com/a/15180",
    "source_tikz": "...",
    "target_tikz": "...",
    "edit_type": "STRUCTURE",
    "edit_operation": 1,
    "license": "CC-BY-SA-2.5"
}

Field Descriptions

Field	Description
id	Unique identifier for each sample.
source_image_file_name	Path to the input diagram.
visual_instruction_file_name	Image-based instruction describing the edit.
target_image_file_name	Path to the reference edited diagram.
uri	Source reference (e.g., original TikZ example).
source_tikz	TikZ code used to generate the source image.
target_tikz	TikZ code after applying the edit.
edit_type	Category of edit (e.g., STRUCTURE, ADD, etc.).
edit_operation	Number of edit operations.
license	CC BY-SA version assigned to the original Stack Exchange contribution using its DaTikZ-v2 date.

Source Licenses

The data is licensed under Creative Commons Attribution-ShareAlike (CC BY-SA). The applicable license version for each sample is recorded in its license field, following the Stack Exchange licensing policy.

Citation

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

VisEditBench: A Benchmark for Vector-Format Diagram Editing with Visual Instructions
Akito Taneguchi, Haruto Yoshida, Jun Suzuki, and Itsumi Saito.
The Fortieth Annual Conference on Neural Information Processing Systems Evaluations and Datasets Track, 2026.

For more information, see the paper on OpenReview and the project page.

@inproceedings{taneguchi2026viseditbench,
  title     = {VisEditBench: A Benchmark for Vector-Format Diagram Editing with Visual Instructions},
  author    = {Akito Taneguchi and Haruto Yoshida and Jun Suzuki and Itsumi Saito},
  booktitle = {The Fortieth Annual Conference on Neural Information Processing Systems Evaluations and Datasets Track},
  year      = {2026},
  url       = {https://openreview.net/forum?id=e4d9v8gOoV}
}