LocateEdit-Bench

A Benchmark for Instruction-Based Editing Localization

LocateEdit-Bench is a large-scale benchmark for localizing the regions modified by instruction-based image editors. Unlike inpainting-based manipulations, where the edited area is given by an explicit mask, instruction-based editors (e.g., MMDiT-based models) produce semantically coherent and visually seamless edits without any predefined region. LocateEdit-Bench provides 231K edited images from four state-of-the-art editors, each paired with its original image, the editing instruction, and a pixel-level ground-truth mask.

Paper: arXiv:2602.05577

Highlights

• 231,092 edited images from four instruction-based editors: Qwen-Image-Edit, Flux-Kontext, Step1X-Edit, and BAGEL.
• Three object-level edit types: object addition, object swap, and attribute modification — edits with a well-defined, object-localized region.
• High-quality masks: target object identified by ChatGPT, segmented by SAM3, filtered at confidence > 0.5.
• 5,504 distinct object types; source images drawn from LAION-5B and OpenImages V6 (via OmniEdit).
• Dedicated train/validation splits per editor: the validation split is the paper's 24K test set (6,000 images per editor, stratified by edit type).

Dataset structure

LocateEdit-Bench/
├── manifest.json      # per-shard split, row-index range, relative path
├── splits.json        # {model: {train: [files], val: [files]}}
├── stats.json         # rows / edit-type counts / sizes per model
├── summary.json       # global totals
├── train/
│   ├── BAGEL/BAGEL_train_00001.parquet ...
│   ├── Flux-Kontext/Flux-Kontext_train_00001.parquet ...
│   ├── Qwen-Image-Edit/Qwen-Image-Edit_train_00001.parquet ...
│   └── Step1X-Edit/Step1X-Edit_train_00001.parquet ...
└── val/
    ├── BAGEL/BAGEL_val_00001.parquet ...
    └── ...

Each parquet file targets ~1 GiB (roughly 370–560 samples). Filenames are <editor>_<split>_<5-digit index>.parquet; numbering restarts at 00001 for each split and each editor. Shards never mix splits, so a file can be assigned to a split by its path alone. splits.json lists every file explicitly.

Schema

One row per edited image. All images are stored as raw PNG bytes (lossless, no re-encoding) with width 1024. The dataset card declares the three image columns as Image, so the Hub viewer and datasets decode them automatically.

Column	Type	Description
img_name	string	Unique sample name / original filename, e.g. addition_100005_4_0.png
model	string	Editing model: BAGEL, Flux-Kontext, Qwen-Image-Edit, Step1X-Edit
split	string	train or val
row_index	int32	0-based position within the split's original sample list
edit_type	string	addition, swap, or attribute_modification
edit_subtype	string	facial for facial attribute edits (BAGEL only), otherwise null
edit_prompt	string	Natural-language editing instruction
sam_prompt	string	Short phrase describing the edited target, used for segmentation
boxes	list<list>	Bounding boxes [x1, y1, x2, y2] of the edited target(s)
scores	list	Segmentation confidence per box
image	binary	Original image, PNG bytes
edited_image	binary	Edited image, PNG bytes
mask	binary	Pixel-level mask of the edited region, PNG bytes

Notes:

• boxes / scores are the raw segmentation outputs; boxes may extend slightly beyond the image boundary.
• Masks are 8-bit PNGs with the same spatial size as their images.

Splits

Editor	Train	Validation	Total
BAGEL	47,320	6,000	53,320
Flux-Kontext	52,707	6,000	58,707
Qwen-Image-Edit	55,681	6,000	61,681
Step1X-Edit	51,384	6,000	57,384
All	207,092	24,000	231,092

Edit-type distribution (all splits):

Editor	Addition	Swap	Attribute Mod.
BAGEL	25,530	16,983	10,807
Flux-Kontext	26,138	18,365	14,204
Qwen-Image-Edit	27,381	18,798	15,502
Step1X-Edit	26,070	17,370	13,944
All	105,119	71,516	54,457

The validation split follows the paper's protocol: 6,000 images are sampled from each editor's outputs, stratified to preserve that editor's edit-type composition.

Construction

1. Source data. Real images and editing instructions are selected from OmniEdit (itself built from LAION-5B and OpenImages V6). Instructions were produced with GPT-4o. Roughly 400K image–instruction pairs were retained after filtering.
2. Edit-type selection. Only object addition, object swap, and attribute modification are kept. Style transfer, background swap, environment change and holistic edits are excluded because they lack a unique object-localized region.
3. Editing. Each edit-type subset is evenly divided among the four editors. All images are resized to width ~1024 before editing; generation follows each model's recommended inference settings.
4. Quality filtering. Edited images are scored for editing accuracy and background consistency; low-confidence samples are discarded.
5. Mask generation. ChatGPT identifies the edited target; SAM3 produces a promptable segmentation for it. Samples with segmentation confidence <= 0.5 are dropped, yielding the final pixel-level masks.

Usage

Option A — datasets (recommended)

from datasets import load_dataset, Image

# One editor (config name = editor name); image columns are decoded automatically
ds = load_dataset("MoeNew/LocateEdit-Bench", "BAGEL")
train, val = ds["train"], ds["validation"]

# All four editors merged (distinguish by the `model` column)
ds_all = load_dataset("MoeNew/LocateEdit-Bench", "all")

# If you want raw PNG bytes instead of decoded PIL images:
raw = ds["train"].cast_column("image", Image(decode=False))

Option B — explicit files (no config), e.g. for local copies

from datasets import load_dataset

ds = load_dataset("parquet", data_files={
    "train": "train/Qwen-Image-Edit/*.parquet",
    "validation": "val/Qwen-Image-Edit/*.parquet",
})

Option C — pyarrow / Parquet directly

import pyarrow.dataset as pads

d = pads.dataset("LocateEdit-Bench", format="parquet", partitioning=None)
tbl = d.to_table(filter=(pads.field("split") == "val"))

Because each shard contains exactly one split (and one editor), filtering on split or model prunes whole row groups without reading image bytes. Note that split/model live in the path rather than as partition columns, so the quickest selection is by file path (splits.json / manifest.json) or by a column filter.

Evaluation protocols

The paper defines two protocols and reports pixel-level and region-level metrics:

• Protocol 1 — Full-set evaluation. Train on the composite training set (all four editors) and test on each editor's validation split.
• Protocol 2 — Cross-editor generalization. Train on a single editor's data and test on every editor, measuring transfer to unseen editing models.
• Metrics. Accuracy, AUC and F1 treat each pixel as a binary decision; Dice and mIoU measure region overlap. Pixel-level accuracy/AUC are dominated by the large untouched background, so Dice/mIoU are the more informative scores.

License and citation

LocateEdit-Bench is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. The source images originate from LAION-5B and OpenImages V6 via OmniEdit, and the edited images were generated with the corresponding editing models; please also respect the terms of those upstream sources and model licenses.

@article{wu2026locateeditbench,
  title   = {LocateEdit-Bench: A Benchmark for Instruction-Based Editing Localization},
  author  = {Shiyu Wu and Shuyan Li and Jing Li and Jing Liu and Yequan Wang},
  journal = {arXiv preprint arXiv:2602.05577},
  year    = {2026}
}