hi-paris/FakeParts_Legacy

FakePartsBench是一个专注于部分深度伪造（partial deepfakes）检测的大规模基准数据集。数据集包含超过25,000个经过处理的视频片段和16,000个真实视频片段，分辨率高达1080p，时长通常在5-14秒之间。数据集提供了像素级和帧级的注释，涵盖了多种类型的深度伪造技术，包括全合成视频（T2V/I2V/TI2V）和部分操纵（如面部交换、修复、外绘、风格变化、插值等）。该数据集旨在解决当前深度伪造检测方法在检测部分视频操纵方面的不足，并提供了训练和评估检测器所需的资源。

FakePartsBench is a large-scale benchmark dataset specifically designed for detecting partial deepfakes. It comprises over 25,000 manipulated video clips and 16,000 real clips, with high-resolution content (up to 1080p) and durations typically ranging from 5 to 14 seconds. The dataset provides pixel-level and frame-level annotations, covering a wide spectrum of deepfake techniques, including fully synthetic videos (T2V/I2V/TI2V) and partial manipulations such as faceswap, inpainting, outpainting, style change, and interpolation. This dataset addresses the critical gap in current deepfake detection capabilities for partial video manipulations and offers the necessary resources for training and evaluating detection methods.