LEIQAD

# LEIQAD: Local Editing Image Quality Assessment Dataset LEIQAD (Local Editing Image Quality Assessment Dataset) is a dataset for image quality assessment in local editing scenarios. It focuses on visual quality changes introduced by region-confined operations, where artifacts are often concentrated around the edited area rather than distributed across the whole image. The dataset supports fine-grained quality modeling, local quality perception, and region-aware image quality assessment. It covers four local quality variation scenarios: - `add`: object or content addition - `remove`: object or content removal - `replace`: object or content replacement - `compression`: localized compression degradation The Add, Remove, and Replace subsets are constructed from AnyEdit-based semantic editing samples. In the paper construction protocol, 8,089 high-quality AI local editing images are retained for these three semantic editing categories after multi-stage filtering. The Compression subset is a complementary local degradation subset and is not included in the 8,089 AnyEdit semantic-editing count. ## Public Release Manifest The public release target is `LEIQAD_use`. The public manifest uses `all_data.csv` as the single metadata entry point and does not include the internal train/test JSON split files. ```text LEIQAD_use/ all_data.csv add/ image_file/ edited_file/ compression/ image_file/ edited_file/ remove/ image_file/ edited_file/ replace/ image_file/ edited_file/ Origin/ add/ image_file/ compression/ image_file/ remove/ image_file/ replace/ image_file/ ``` The task directories contain original images in `image_file/` and edited or degraded images in `edited_file/`. The `Origin/` directory provides a convenience collection of unique original images grouped by task type. ## Metadata Fields `all_data.csv` contains one row per released assessment sample. | Field | Description | |---|---| | `edit_instruction` | Text instruction or degradation description. | | `edit_type` | One of `add`, `remove`, `replace`, or `compression`. | | `image_file_path` | Relative path to the original image within the corresponding task directory. | | `edited_file_path` | Relative path to the edited/degraded image within the corresponding task directory. | | `mos` | Final released quality score synchronized from the normalized JSON score. | Example row layout: ```text edit_instruction,edit_type,image_file_path,edited_file_path,mos ``` For a row with `edit_type=add`, the corresponding files are: ```text LEIQAD_use/add/<image_file_path> LEIQAD_use/add/<edited_file_path> ``` ## Release Statistics Current public release: | Item | Count | |---|---:| | `all_data.csv` rows | 11,008 | | Published metadata files | 1 | | Published image files | 25,772 | | Published checksum entries | 25,773 | Task directory image counts: | Subset | Image Files | |---|---:| | add | 5,206 | | compression | 3,826 | | remove | 4,548 | | replace | 4,810 | | Total | 18,390 | `Origin/` image counts: | Subset | Image Files | |---|---:| | add | 2,603 | | compression | 100 | | remove | 2,274 | | replace | 2,405 | | Total | 7,382 | The total published image count is 25,772, including task-directory images and the `Origin/` convenience image collection. ## Loading Example ```python import csv from pathlib import Path root = Path("LEIQAD_use") with (root / "all_data.csv").open("r", encoding="utf-8", newline="") as f: rows = list(csv.DictReader(f)) sample = rows[0] subset = sample["edit_type"] original = root / subset / sample["image_file_path"] edited = root / subset / sample["edited_file_path"] score = float(sample["mos"]) print(original, edited, score) ``` ## Dataset Construction Summary LEIQAD adopts a local-first annotation paradigm. Unlike conventional IQA datasets that primarily rely on global MOS, LEIQAD centers subjective assessment on the edited region and its surrounding context. The semantic editing subsets are based on AnyEdit-derived Add, Remove, and Replace samples. Samples are filtered through automatic screening, manual review, and subjective representativeness review to ensure editing locality and practical relevance. The subjective experiment follows a double stimulus continuous quality scale (DSCQS) protocol. Participants evaluate edited results with reference to the original image and text instruction. Scores are guided by three dimensions: - Instruction Fidelity - Local Quality - Regional Harmony After quality control, ratings from valid participants are used to compute MOS and normalized quality scores. The public `mos` field in `all_data.csv` stores the final released quality score. ## Integrity Check `checksums.txt` contains SHA-256 hashes for the public release files. It covers `LEIQAD_use/all_data.csv`, all released image files, and the `Origin/` image collection. It intentionally excludes internal train/test JSON split files. To verify a file manually: ```powershell Get-FileHash .\LEIQAD_use\all_data.csv -Algorithm SHA256 ``` ## License and Usage This dataset is released under a restricted research license. It is intended for non-commercial research and educational use only. The underlying AnyEdit samples, source images, edited images, and any third-party assets may remain subject to their original licenses or terms. See `LICENSE` for details. Before redistributing or using the dataset beyond non-commercial research, users are responsible for confirming that all upstream rights and licenses permit the intended use. ## Citation Citation information will be updated after the associated paper is finalized. ```bibtex @dataset{leiqad, title = {LEIQAD: Local Editing Image Quality Assessment Dataset}, author = {Author Name(s)}, year = {2026}, note = {Dataset for local editing image quality assessment} } ```