EditableDesignBench

Code and evaluation instructions · Download ZIP

EditableDesignBench contains 600 image-editing tasks for fixed-canvas visual designs. It has two tracks:

Track	Tasks	Source designs	Operations
Text (default or text)	400	200	Delete, insert, replace, move, scale, recolor
Image assets (asset)	200	34	Delete, insert, replace, move, scale

The asset track contains 40 editable asset units with five tasks per unit, using 39 distinct generated replacement materials. Its source designs are a subset of the same 200 original designs; 600 tasks are not 600 independent source images. All configurations use the test split. Group by source_id across both tracks when splitting data or estimating uncertainty.

Download and run

The evaluation repository documents installation, API configuration, model generation and scoring. Its dataset downloader installs both tracks in one command:

python scripts/download_dataset.py

For an existing older installation, use python scripts/download_dataset.py --update. The previous local directory is backed up after the new download passes validation.

The ZIP has one top-level EditableDesignBench/ directory. Text tasks remain at the root for compatibility; asset tasks are in assets/.

Alternatively, load the Parquet configurations:

from datasets import load_dataset

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

Input and reference data

Each editing request receives exactly one source image and its original natural-language instruction, independently. Reference targets, masks and other tasks are not model inputs.

Each track has tasks.jsonl, references.jsonl and cases/<id>/. A case contains the native-size RGB source.png and target.png, binary masks, and an annotation describing its operation and geometry. HTML source files, credentials and model outputs are not included.

The asset track includes allowed_edit_mask.png: white pixels denote the source/instruction-defined editing allowance excluded from background-preservation metrics. references.jsonl identifies it as pixel_preservation_mask. The exact reference source/target union (eval_mask.png) is retained for construction inspection and is not the asset background-scoring mask.

Asset insertion sources have the original asset removed. Inserted and replacement reference assets were created with the built-in image_gen tool and composed into the design. Images and their inseparable frames or rings are edited together when the instruction specifies that unit. Many different inserted or replacement images can satisfy an instruction; reproducing the one reference bitmap is not required.

Evaluation

The code reports MSE, PSNR, SSIM and LPIPS for background preservation using the TextEditBench implementations. The text track uses its source/target edit-mask union. The asset track uses its source/instruction allowance.

The MLLM judge makes four independent 0–5 scoring requests per task:

Track	Criteria
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)

Asset judges receive the full source and prediction plus matching source-defined detail views, the original instruction and the operation. They do not receive the reference target, hidden replacement material, or the editing model's name. Natural movement/scale variations are accepted where the instruction leaves them unspecified. Unrelated text and graphics remain protected, including inside rectangular editing regions.

Asset results include per-dimension and per-operation scores and a joint qualification rate: all four dimensions must score at least 4. Its denominator is all selected tasks; missing model outputs count as unqualified. API failures remain pending and are retried rather than converted into model scores. Scores and coverage are reported together. Asset criteria and the joint rate are project-specific extensions, not an official TextEditBench asset protocol. Report the two tracks separately.

Release 2.0.0

This release updates the existing 400 text tasks to the locally validated text revision 2, correcting 20 target layouts that could overlap or clip neighboring content. Task IDs are retained and source PNGs are unchanged; three movement instructions were updated together with their references. It adds the 200-task asset track, with 18 enlargements and 22 reductions among the scaling tasks. Re-run changed instructions rather than treating predictions generated for the old task as current outputs.

Construction checks cover editable-unit identity, clipping, text/graphic collisions, rectangular mask coverage, and unchanged pixels outside the editing region. The archive includes published inputs and references rather than local construction logs or browser audit files.

License and citation

The dataset is released under CC BY-NC 4.0. The evaluation code is licensed under MIT. Please also acknowledge the upstream work listed in the code repository's notices.

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