GenImage
收藏资源简介:
GenImage数据集是用于AI生成图像检测的大型多元化数据集,旨在为研究社区提供一个统一的基准,以评估检测方法。该数据集基于ImageNet数据集,包含了来自ImageNet的自然图像以及由八种不同生成模型产生的AI生成图像,确保了自然图像与AI生成图像在内容分布上的一致性。数据集分为八个不同的子集,每个子集包含训练和验证数据,用于评估检测方法的跨生成器性能。GenImage数据集的关键在于揭示了现有数据集在JPEG压缩和图像尺寸方面的偏差,这些偏差影响了检测器的效能和评估。通过移除这些偏差,数据集显著提高了检测器对JPEG压缩的鲁棒性和跨生成器性能,为生成图像检测领域提供了更为准确和可靠的评估基准。
The GenImage dataset is a large-scale, diverse dataset for AI-generated image detection, aiming to provide the research community with a unified benchmark for evaluating detection methods. Built upon the ImageNet dataset, this collection includes natural images sourced from ImageNet and AI-generated images produced by eight distinct generative models, ensuring consistency in content distribution between natural and AI-generated images. The dataset is divided into eight distinct subsets, each containing training and validation data for evaluating the cross-generator performance of detection methods. A core aspect of the GenImage dataset is its revelation of biases in existing datasets related to JPEG compression and image size, which impair the effectiveness and evaluation of detectors. By eliminating these biases, the dataset significantly enhances the robustness of detectors against JPEG compression and their cross-generator performance, providing a more accurate and reliable evaluation benchmark for the field of generated image detection.

- 1Fake or JPEG? Revealing Common Biases in Generated Image Detection Datasets弗劳恩霍夫ITWM,高性能计算能力中心凯撒斯劳滕,德国 2分析机器学习和分析研究所(IMLA),奥芬堡大学,德国 3曼海姆大学,德国 · 2024年



