DUET_v1 — Image Dataset

Project: DUET_v1
Course: SC4001 — Nanyang Technological University
Academic Year: 2026/27
Team: Group 66
Organization: DUET-Bench

Overview

This repository hosts curated image datasets used in DUET_v1, a project developed by Group 66 for NTU SC4001.

The datasets support the development and evaluation of image authenticity detection methods, with an emphasis on reproducible data preparation, structured storage, and sample-level traceability.

Dataset Collection

NTIRE / Fake

The ntire_fake configuration contains 50,000 images labeled as fake, stored under NTIRE/Fake/.

The images are organized into self-contained Parquet shards, with the original JPEG bytes embedded in the image column.

All selected images satisfy the following resolution requirements:

• Width > 512 pixels
• Height > 512 pixels

After applying the dimension filter, deterministic sample selection preserves the original ordering of each source labels.csv.

Dataset composition:

Source Shard	Number of Images
shard_5	13,024
shard_1	23,732
shard_2	13,244
Total	50,000

Data Integrity and Provenance

To support reproducibility and data verification, the repository includes:

• NTIRE/Fake/selection_manifest.csv — Records sample-level provenance and SHA-256 hashes.
• NTIRE/Fake/summary.json — Documents dataset construction and verification details.

Usage

The dataset can be accessed through the Hugging Face Datasets library.

from datasets import load_dataset

dataset = load_dataset(
    "DUET-Bench/Dataset",
    "ntire_fake",
    split="train"
)

Repository: DUET-Bench/Dataset

Usage and Licensing

This repository is intended for research and educational use.

Before redistributing images or using the dataset beyond the permitted scope, users should verify the licenses, attribution requirements, and usage restrictions of the original source datasets.

Project

For more information about the DUET_v1 project and associated resources, visit the DUET-Bench Hugging Face organization.