EditableDesignBench / README.md
yomi-chen's picture
Release v4.0.0: 100 text and 100 asset editing tasks
6b2b925 verified
|
Raw History Blame Contribute Delete
5.3 kB
metadata
pretty_name: EditableDesignBench
license: cc-by-nc-4.0
language:
  - en
task_categories:
  - image-to-image
size_categories:
  - n<1K
tags:
  - image-editing
  - text-editing
  - asset-editing
  - graphic-design
  - synthetic-data
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*.parquet
  - config_name: text
    data_files:
      - split: test
        path: data/test-*.parquet
  - config_name: asset
    data_files:
      - split: test
        path: assets/data/test-*.parquet

EditableDesignBench

Code and evaluation · Download ZIP

EditableDesignBench contains 200 image-editing tasks for fixed-canvas designs.

Track Tasks Source designs Operations
Text (default or text) 100 100 Delete, insert, replace, move, scale, recolor
Image assets (asset) 100 34 Delete, insert, replace, move, scale

The collection uses 100 distinct original designs. Every asset source is also represented in the text track. Asset tasks cover all 40 editable asset units from the parent release, with two or three operations per unit. Tasks sharing a source_id are related observations, not independent source images.

Download and run

Follow the evaluation repository for environment installation, API configuration and commands. Download both tracks with:

python scripts/download_dataset.py

The current release is installed into data/EditableDesignBench-v4/. The ZIP contains one top-level EditableDesignBench/ directory: text tasks at the root and asset tasks in assets/.

Alternatively:

from datasets import load_dataset

text = load_dataset("yomi-chen/EditableDesignBench", "text", split="test")
assets = load_dataset("yomi-chen/EditableDesignBench", "asset", split="test")

Model inputs and reference files

Each model receives one source image and the original natural-language instruction, independently for each task. Reference targets, masks, HTML, other tasks and model outputs are not model inputs.

Each track provides tasks.jsonl, references.jsonl, binary rectangular masks, native RGB PNGs and annotations in cases/<id>/. The asset pixel_preservation_mask points to allowed_edit_mask.png, the source/instruction-defined editing allowance. Its eval_mask.png records the particular reference change for construction inspection and is not the asset background-scoring mask. Inserted or replacement reference images are not supplied to the editing model, and reproducing their exact pixels is not required.

Evaluation

Pixel preservation uses TextEditBench's MSE, PSNR, SSIM and LPIPS implementations. The MLLM judge receives the full source and candidate images and the instruction, and one API request returns all four 0–5 scores:

Track Scores
Text Instruction following (IF), text accuracy (TA), visual consistency (VC), layout preservation (LP)
Assets Instruction following (IF), asset-state accuracy (OA), visual coherence (VC), non-target layout/content preservation (LP)

There are 200 judge requests per model for the complete collection, excluding retries. No mask, reference target, detail crop or model identity is supplied to the default judge. Temperature is 0 and the standard output budget is 500 tokens. Short per-criterion reasons accompany the scores.

Successful scores are cached. Network/API failures and actually missing or invalid scores can be retried; missing reasons or auxiliary evidence and score/evidence disagreement are diagnostic only. Valid scores are not rejected or rescaled. Joint qualification requires all four scores to be at least 4. Report coverage and the two tracks separately. These asset criteria and the joint-request adaptation are project extensions, not an unchanged official TextEditBench scoring protocol.

Release 4.0.0

This is a selection-only data revision from v3.0.0: 100 text and 100 asset tasks. Selected instructions, PNGs, masks and annotations are byte-for-byte unchanged. Selection uses seed 42 and only coverage metadata—never model predictions or evaluation scores.

  • Text: one task per source, 20 per design domain, and 16–17 per edit operation. All asset source designs are retained.
  • Assets: 20 per operation, every parent source and asset unit retained, two or three tasks per asset unit. Repeated use of source 134 is necessary because it has five different asset units.
  • Exact selections and policies are in text_subset.json and assets/asset_subset.json.

Group by source_id across both tracks when creating splits or uncertainty estimates. Older data remain available at revision v3.0.0. Aggregates over different selections and scores from joint versus independent requests must be identified separately.

License and citation

Data: CC BY-NC 4.0. Evaluation code: MIT. See NOTICE.md and licenses/ for upstream attribution.

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