license: apache-2.0
task_categories:
- image-feature-extraction
- feature-extraction
language:
- en
tags:
- multimodal
- celebrity
- attribute
- image-text
This repository contains the Celeb Attributes Portfolio (CAP) dataset, which is part of the research presented in the paper Beyond Artificial Misalignment: Detecting and Grounding Semantic-Coordinated Multimodal Manipulations.
Code: https://github.com/shen8424/SAMM-RamDG-CAP
Introduction
We present CAP, a large-scale database including over 70k celebrities. Each celebrity in the CAP has at least three associated images along with their gender, birth year, occupation, and main achievements.
CAP serves as an auxiliary dataset for the Semantic-Aligned Multimodal Manipulation (SAMM) dataset, providing contextual evidence for the Retrieval-Augmented Manipulation Detection and Grounding (RamDG) framework, as described in our paper. It enhances the realism of multimodal manipulation detection by offering semantically consistent visual and textual attributes for grounding manipulations.
Two examples from CAP:
We provide the Python script import_cap.py that automatically integrates CAP information into your custom dataset by identifying all public celebrities mentioned in the textual content of each entry and incorporating their corresponding information.
Usage 🔧
If you want to import the CAP data into your own dataset, please refer to this.
🤗🤗🤗 Citation
If you find this work useful for your research, please kindly cite our paper:
@inproceedings{shen2025beyond,
title={Beyond Artificial Misalignment: Detecting and Grounding Semantic-Coordinated Multimodal Manipulations},
author={Shen, Jinjie and Wang, Yaxiong and Chen, Lechao and Nan, Pu and Zhong, Zhun},
booktitle={Proceedings of the ACM International Conference on Multimedia (MM)},
year={2025},
url={https://huggingface.co/papers/2509.12653}
}