madebyollin/movirec

MoViRec是一个用于评估小型视频VAE的小型视频重建基准数据集。它包含500个256x256分辨率、5秒时长的视频片段，这些片段来自Reddit上的AI生成视频。数据集还包括一个视频特征提取器（V-JEPA + JEDi SSv2微调的池化块）和一个评估笔记本（MoViRec.ipynb），用于测量视频VAE的特征重建误差。数据集的目的是为视频VAE的实际应用提供一个快速、可信的评估工具。视频片段经过下采样和帧率调整以减少高频伪影，同时偏向于AI生成视频的内容。MoViRec is a small video reconstruction benchmark for benchmarking small video VAEs. It consists of 500 256x256 5-second video patches drawn from various AI video subreddits, a video featurizer (V-JEPA + the JEDi SSv2 fine-tuned pooling block), and an evaluation notebook (MoViRec.ipynb) which measures the feature reconstruction error of each VAE. The dataset is designed to provide a fast and trustworthy way to evaluate the performance of video VAEs, particularly for AI-generated video content. The video patches are processed to minimize high-frequency artifacts by downscaling and subsampling, with a focus on reflecting the actual use cases of video VAEs.