EditableDesignBench

EditableDesignBench is a benchmark for instruction-guided text editing in designed images. It contains 400 editing tasks from 200 source designs, spanning six editing operations and five document types.

Each task provides an RGB source image, a short natural-language instruction, an RGB target image, rectangular region masks, and structured annotations. The default model input is the source image and instruction; target images and evaluation annotations are used after generation.

Quick start

pip install datasets pillow

from datasets import load_dataset

dataset = load_dataset("yomi-chen/EditableDesignBench", split="test")
sample = dataset[0]
source_image = sample["source_image"]
instruction = sample["instruction"]
print(sample["id"], instruction)

The Parquet files embed all images and masks, so the dataset can be loaded directly without extracting the archive.

For the original PNG/JSON file layout:

pip install huggingface_hub
hf download yomi-chen/EditableDesignBench EditableDesignBench.zip --repo-type dataset --local-dir .
unzip EditableDesignBench.zip

import json
from pathlib import Path
from PIL import Image

root = Path("EditableDesignBench")
tasks = [json.loads(line) for line in (root / "tasks.jsonl").read_text().splitlines()]
source_image = Image.open(root / tasks[0]["image"])
instruction = tasks[0]["instruction"]

Dataset composition

All 400 cases are provided in one test split. Each source design contributes two tasks. When constructing additional splits, group by source_id / group_id to keep tasks from the same design together.

Editing operation	Cases
Delete	69
Insert	69
Replace	55
Move	69
Scale	69
Change color	69

Each of the five document types has 80 cases: webpages, slides, posters, scientific figures, and charts. Images retain their native resolutions; there is no mandatory square-image resize.

Fields

Field	Meaning
id	Stable editing-task identifier
source_id, group_id	Source-design identity used for grouping
domain	Document type
original_dimension	Capability dimension of the original design prompt
operation	Editing operation
instruction	Short natural-language command
source_image	RGB input image
target_image	RGB reference result
source_mask	Editing region in the source image
target_region_mask	Editing region in the target image
eval_mask	Union of source and target regions
width, height	Native image dimensions
annotation_json	JSON string with text, geometry, style, editing parameters, and construction checks

Masks use 255 for the editing region and 0 for the remaining area. They are unions of rectangular quadrilaterals derived from rendered text boxes with an 8-pixel margin, rather than character-contour masks. Coordinates use the native image canvas with the origin at the top left. For move operations, the source mask marks the original position, while the evaluation mask also covers the destination.

The PNG/JSON archive has this layout:

EditableDesignBench/
  tasks.jsonl
  references.jsonl
  inventory.csv
  summary.json
  validation.jsonl
  cases/<id>/
    source.png
    target.png
    mask.png
    target_region_mask.png
    eval_mask.png
    task.json
    annotation.json

tasks.jsonl is the model-input manifest. references.jsonl links the target images and evaluation annotations by task ID.

Construction and evaluation

Image pairs and region annotations are constructed from controlled edits to rendered HTML/SVG designs. The delivered PNGs are the evaluation reference. The release contains image pairs, masks, instructions, and annotations.

Construction checks cover image dimensions and modes, binary masks, region unions, and localization of rendered changes. The data were constructed automatically with representative visual review; exhaustive independent human semantic verification is not claimed.

The evaluation convention follows TextEditBench: background-preservation metrics (MSE, PSNR, SSIM, LPIPS) use the supplied regions, while a multimodal judge can assess instruction following (IF), text accuracy (TA), visual consistency (VC), and layout preservation (LP). A model result must be evaluated against the original instruction. Do not provide the reference target image or evaluation masks to an instruction-only editing model.

License and acknowledgment

The dataset is released under CC BY-NC 4.0: attribution is required and use is limited to noncommercial purposes. See LICENSE and NOTICE.md.

The original design prompts are based on BizGenEval. The editing-benchmark setup is inspired by TextEditBench. EditableDesignBench is a separate dataset, with newly constructed editing pairs and annotations.

Citation

@misc{editabledesignbench2026,
  title = {EditableDesignBench},
  author = {{yomi-chen}},
  year = {2026},
  url = {https://huggingface.co/datasets/yomi-chen/EditableDesignBench}
}