COCO/README.md

PAPE — COCO

Real COCO photographs edited by four image-editing models. Each manifest row is one language-specific pair: the source image, a semantically normal replacement, a semantically abnormal replacement of the same object, and a binary mask for each edit. Every pair in this folder has normal/abnormal mask IoU of at least 0.5.

2,000 distinct COCO image ids are included. A source image can appear more than once when several objects, anomaly categories, or languages were edited.

Counts

Editor	Directory	Pairs	Chinese	English
Qwen-Image-Edit-2511	qwen_image_edit_2511	3,754	1,864	1,890
FireRed-Image-Edit-1.1	firered_image_edit_1.1	3,832	1,915	1,917
LongCat-Image-Edit	longcat_image_edit	3,511	1,706	1,805
FLUX.2-dev	flux2_dev	3,571	1,744	1,827
Total		14,668	7,229	7,439

Split

splits/split_seed2026.json assigns COCO image_id values with seed 2026: 1,600 train, 200 validation, and 200 test. The same image id is used for every editor. ImageNet and Places are held out as test.

Files

<edit_model>/
  manifest.jsonl
  source/          JPEG, suffix .jpg
  normal/          PNG
  abnormal/        PNG
  masks/normal/    PNG
  masks/abnormal/  PNG

Filenames follow {sample_id}__{language}_{role}, with roles source, normal, abnormal, normal_mask, and abnormal_mask.

manifest.jsonl fields used to load a row:

Field	Meaning
sample_id	Source image, target object, and anomaly id
image_id	COCO file id, also the split key
language	zh or en
edit_model_name	Editor directory name
anomaly_type	One category below
iou	IoU between the normal and abnormal masks
normal_prompt, abnormal_prompt	Instructions applied to produce this row
normal_prompt_zh, normal_prompt_en, abnormal_prompt_zh, abnormal_prompt_en	Bilingual prompt fields stored with the row

source_paths and export_paths are absolute paths from the original export machine. Load images from source/, normal/, abnormal/, and masks/ in this folder.

Anomaly categories

Id	Pairs	Share	Abnormal edit
C1_physical_attribute	3,566	24.3%	Material or physical state is not one the object can have
C2_spatial_relation	1,965	13.4%	Support, contact, containment, or scale is inconsistent
C3_function_affordance	2,477	16.9%	The object does not support the use shown in the edit
C4_scene_context	2,008	13.7%	A recognizable object does not belong in the scene
C5_biological_structure	2,392	16.3%	Anatomy, part count, or bodily configuration is inconsistent
C6_interaction_relation	2,260	15.4%	An action or relation between entities is incoherent

The normal edit replaces the same object in a semantically ordinary way. Both edits are meant to keep location, scale, lighting, and viewpoint close to each other.

COCO photographs remain under the COCO dataset terms. This folder does not relicense them.