Figureure 2 · : Left: Distribution of videos across hierarchical evaluation levels aFig. 2: Left: Distribution of videos across hierarchical evaluation levels along with the proportion of samples subjected to controlled distortions. Right: Illustration of the controlled distortion pipeline. High-quality videos are first segmented into 15-second clips. Controlled distortions are then applied according to predefined distortion pools and configurations (LQU, CQR, GQU). Finally, distorted clips are merged into fulllength videos to enable distortion-aware question-answer generation.这张图概括 LongVQUBench 的整体方法流程。阅读时先看模块之间传递的训练信号,再看作者如何把目标拆成可优化的子问题。FigFig. S9: Illustration of saturation shift across diferent intensity levels.这张图来自论文 PDF 的结构化抽取。当前用于辅助理解 LongVQUBench 的方法或实验,请结合正文精读段落一起看。