OmniDiff

OmniDiff is a benchmark for fine-grained Image Difference Captioning (IDC): given two similar images, a model must describe their subtle visual differences in natural language. The dataset combines complex real-world environments with near-realistic 3D-rendered scenes and emphasizes precise change localization and detailed semantic descriptions.

The dataset is introduced in “OmniDiff: A Comprehensive Benchmark for Fine-grained Image Difference Captioning” by Yuan Liu, Saihui Hou, Saijie Hou, Jiabao Du, Shibei Meng, and Yongzhen Huang (ICCV 2025).

Highlights

• 15,598 image pairs and captions from 324 scenes.
• 8,609 real-world pairs and 6,989 rendered pairs.
• 12 change types spanning semantic changes and visual distractors.
• Human-written, fine-grained captions averaging 60 words.
• Indoor and outdoor scenes split into train, validation, and test sets at an approximately 8:1:1 ratio.
• One centralized JSON Lines file, annotation.jsonl, for every subset and split.

Dataset construction

The paper describes a two-stage construction pipeline.

1. Change-pair collection

The Real subset covers 224 scenes collected through on-site photography and carefully selected online videos. Cameras were placed in varied indoor and outdoor environments, and similar frames with subtle changes were selected.

The Render subset covers 100 scenes: 50 indoor and 50 outdoor environments. High-quality 3D assets were rendered with Blender, and scene objects were manually modified to create realistic changes beyond simple tabletop settings.

2. Difference-caption collection

Captions were written by human annotators rather than generated automatically. The annotation guidelines in the paper ask annotators to:

1. Describe meaningful semantic changes while ignoring irrelevant visual variation when it is only a distractor.
2. Identify the changed entity with sufficient attributes and spatial context.
3. State how that entity changed, including its before and after state when applicable.

The captions were subsequently reviewed for spelling and grammar. A pair may contain one or more of the following change types:

viewpoint, illumination, addition, disappearance, removal, substitution, size, color, orientation, pose, ocr, and counting.

Statistics

Release files

Subset	Split	Unique images	Image pairs
Real	train	1,674	7,007
Real	val	189	799
Real	test	191	803
Render	train	975	5,445
Render	val	135	777
Render	test	128	767
Total		3,292	15,598

Caption statistics reported in the paper

Statistic	Value
Captions	15,598
Average words per caption	60.0
Total sentences	38,653
Average sentences per caption	2.5
Vocabulary size	3,793
Captions longer than 20 words	37%

Directory layout

OmniDiff/
├── annotation.jsonl
├── README.md
├── Real/
│   ├── train/images/
│   ├── val/images/
│   └── test/images/
└── Render/
    ├── train/images/
    ├── val/images/
    └── test/images/

All paths in annotation.jsonl are relative to the OmniDiff root. Image names are ASCII-only and zero-padded. Rendered filenames additionally indicate whether the source scene is indoor or outdoor. Extensions match the actual JPEG or PNG encoding.

Unified annotation format

annotation.jsonl is the sole source of ground-truth annotations. It uses JSON Lines: each non-empty line is one independent JSON object, allowing streaming reads and direct ingestion by dataset tools. Each line has the following form:

{
  "id": "real_train_pair_000001",
  "subset": "real",
  "split": "train",
  "image": [
    "Real/train/images/real_train_000019.jpg",
    "Real/train/images/real_train_000009.jpg"
  ],
  "caption": "A fine-grained description of the observed differences."
}

Field definitions:

Field	Type	Description
id	string	Globally unique, zero-padded annotation identifier.
subset	string	real or render.
split	string	train, val, or test.
image	array[string]	Ordered pair: first image followed by second image.
caption	string	Human-written ground-truth difference description.

IDs follow <subset>_<split>_pair_<six-digit index>, for example render_test_pair_000001.

Loading annotations

import json
from pathlib import Path

root = Path("/path/to/OmniDiff")
with (root / "annotation.jsonl").open(encoding="utf-8") as handle:
    annotations = [json.loads(line) for line in handle if line.strip()]

train = [item for item in annotations if item["split"] == "train"]

first_image = root / train[0]["image"][0]
second_image = root / train[0]["image"][1]
caption = train[0]["caption"]

Evaluation

The paper evaluates change captioning with BLEU-4, METEOR, ROUGE-L, CIDEr-D, and SPICE using the Microsoft COCO caption-evaluation protocol. OmniDiff-Real and OmniDiff-Render are evaluated separately in a multi-change captioning setting.

Citation

@inproceedings{liu2025omnidiff,
  title     = {OmniDiff: A Comprehensive Benchmark for Fine-grained Image Difference Captioning},
  author    = {Liu, Yuan and Hou, Saihui and Hou, Saijie and Du, Jiabao and Meng, Shibei and Huang, Yongzhen},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  pages     = {21440--21449},
  year      = {2025}
}

License

OmniDiff is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). You may share and adapt the dataset for any purpose, including commercial use, provided that you give appropriate credit, link to the license, and indicate whether changes were made. See LICENSE for the license notice and the official legal code.