PatrickZZZ/PAPE_Dataset

PAPE

PAPE (Plausible Anomalous Paired Edits) is a paired image-editing dataset for localizing semantic edits. Each sample edits one object in a real photograph twice: a semantically ordinary replacement, and a replacement that introduces one controlled semantic anomaly. Both edits have a binary change mask. Only pairs whose normal and abnormal masks have IoU of at least 0.5 are included.

Source photographs come from COCO, ImageNet, and Places. Those photographs stay under the terms of the source dataset. This repository does not relicense them. Generated edits and masks are released for research.

Subsets

Subset	Source images	Language-specific pairs	Role in splits/split_seed2026.json	Readme
COCO	2,000	14,668	train 1,600 / val 200 / test 200 image ids	COCO/README.md
ImageNet	2,500	18,488	all 2,500 image ids are test	ImageNet/README.md
Places	2,000	14,591	all 2,000 image ids are test	Places/README.md
Total	6,500	47,747	seed 2026

A pair is one language. Chinese and English edits of the same object are separate rows, so one source image can contribute more than one pair, and not every image has both languages for every editor.

Editors

Each subset contains the same four edit models:

Directory	Model
qwen_image_edit_2511	Qwen-Image-Edit-2511
firered_image_edit_1.1	FireRed-Image-Edit-1.1
longcat_image_edit	LongCat-Image-Edit
flux2_dev	FLUX.2-dev

Layout

<COCO|ImageNet|Places>/<edit_model>/
  manifest.jsonl
  source/
  normal/
  abnormal/
  masks/normal/
  masks/abnormal/

manifest.jsonl has one JSON object per line. normal_prompt and abnormal_prompt are the instructions used for that row. image_id joins a row to splits/split_seed2026.json. source_paths and export_paths record paths on the machine that built the release; use the files in this repository instead.

Details, counts, and anomaly categories are in each subset readme.