IVC-liuyuan/training_m3diff

M³Diff Instruction-Tuning Data

This repository contains the instruction-tuning annotations used to train M³Diff, the model introduced in OmniDiff: A Comprehensive Benchmark for Fine-grained Image Difference Captioning (ICCV 2025).

The dataset contains 896,015 question-answer records for fine-grained Image Difference Captioning (IDC). Each record presents two related images and asks the model to describe the visual changes between them. This repository contains the JSON annotations and relative image paths only; the image files are not included.

Hugging Face dataset ID: IVC-liuyuan/training_m3diff.

Dataset composition

The training mixture combines OmniDiff with several established IDC datasets covering real-world, edited, surveillance, bird, and 3D-rendered scenes.

Source	Instruction records
CLEVR-Change	444,156
CLEVR-DC	384,466
OmniDiff	28,056
spot-the-diff	20,862
birds-to-words	14,265
IEdit	4,210
Total	896,015

The counts above refer to instruction records, not necessarily unique image pairs.

OmniDiff itself is a fine-grained IDC benchmark with 15,598 human-captioned image pairs collected from 324 diverse indoor and outdoor scenarios. It contains real-world and 3D-rendered scenes, covers 12 visual change types, and has captions averaging 60 words. The change taxonomy includes viewpoint, illumination, addition, disappearance, removal, substitution, size, color, orientation, pose, OCR, and counting.

Data format

training_m3diff.json is one JSON array. Every record has the following structure:

{
  "image_id": "005422.png",
  "image": [
    "CLEVR-Change/nsc_images/CLEVR_nonsemantic_005422.png",
    "CLEVR-Change/sc_images/CLEVR_semantic_005422.png"
  ],
  "conversations": [
    {
      "from": "human",
      "value": "What modifications can be observed between these two pictures? <image><image>Difference:"
    },
    {
      "from": "gpt",
      "value": "the small yellow cylinder changed to brown."
    }
  ]
}

Fields:

• image_id or id: source-specific sample identifier. Older converted records use image_id, while some records use id.
• image: two paths relative to the common image root, ordered as the first and second images being compared.
• conversations: a two-turn LLaVA-style conversation. The human turn contains two <image> placeholders and the GPT turn contains the target difference caption.

All 896,015 records contain exactly two image paths and one human/GPT conversation pair. The data uses 30 paraphrased prompt templates to reduce dependence on a single question formulation.

Loading

Load the annotations directly from the Hub:

from datasets import load_dataset

dataset = load_dataset("IVC-liuyuan/training_m3diff", split="train")
print(dataset[0])

Alternatively, download the raw JSON:

from huggingface_hub import hf_hub_download

annotation_path = hf_hub_download(
    repo_id="IVC-liuyuan/training_m3diff",
    repo_type="dataset",
    filename="training_m3diff.json",
)

To train M³Diff, arrange the separately obtained images under a common root so the paths in each record resolve as follows:

<IMAGE_ROOT>/
├── OmniDiff/
├── CLEVR-Change/
├── CLEVR-DC/
├── IEdit/
├── spot-the-diff/
└── birds-to-words/

Intended use

This dataset is intended for research on image difference captioning, fine-grained paired-image understanding, and multimodal instruction tuning. It was assembled for training M³Diff, which augments LLaVA-OneVision-7B with a Multi-scale Differential Perception (MDP) module.

License

The repository aggregates annotation records derived from multiple datasets, so no single license is asserted for the complete mixture. The annotations, image paths, and separately downloaded images remain subject to the terms of OmniDiff and their respective upstream datasets: Spot-the-Diff, IEdit, Birds-to-Words, CLEVR-Change, and CLEVR-DC. Consult each source before use or redistribution. Uploading this JSON file does not relicense any image.

Citation

If this training set or M³Diff is useful in your research, please cite:

@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},
  pages={21440--21449},
  year={2025}
}

Acknowledgements

M³Diff builds on OmniDiff, LLaVA-OneVision, Qwen2, and SigLIP. We also acknowledge the creators of Spot-the-Diff, IEdit, Birds-to-Words, CLEVR-Change, and CLEVR-DC. Please cite the corresponding upstream works when using their portions of this training mixture.