Dataset

ConsumerGenEdit Dataset

Real on-device generative edits from Apple, Samsung and Google devices, each paired with its original and three learned-compression variants.

Demo subset

Download

4,000 original–edited pairs in eight groups (Apple, Galaxy, Pixel and five cross-device chains), 500 per group, each with Q6, Q4 and Q2 compressed versions, plus metadata and per-operation manifests.

Download demo subset ConsumerGenEdit-demo.zip · about 6 GB · 31,000 files
Full benchmark

Request access

4,000 source images, 48,000 edits and 144,000 compressed variants, with captions in English, Bangla and Farsi, bounding boxes and human references. Shared with researchers on request: email Dr. Abdur R. Shahid with your name, affiliation and intended use.

shahid@cs.siu.edu

Layout

ConsumerGenEdit-demo/
├── README.md
├── metadata.json                  one entry per pair: group, platform, device, category, files
├── remove.json  move.json  generate.json
│                                  manifests by edit operation, with paths to all variants
├── apple/  galaxy/  pixel/        single-device groups, 500 pairs each
└── crossdevice_<source>_<editor>/ five cross-device groups, 500 pairs each
    ├── original/                  <category>_<id>_original.jpg
    ├── edited/                    <category>_<id>_edited.jpg
    └── compressed/Q6 Q4 Q2/       same file names, original and edited

metadata.json entry
{"id": "000", "group": "apple", "platform": "apple", "device": "apple_ipad",
 "is_crossdevice": false, "category": "airport", "scene_key": "airport::000",
 "original_file": "airport_000_original.jpg", "edited_file": "airport_000_edited.jpg",
 "compressed_available": {"Q2_original": true, "Q2_edited": true, ...}}

Non-commercial research use only. Images follow the terms of the editing platforms and source-image licenses; please do not redistribute, and cite the paper if you use the data.