Generating Videos with Scene Dynamics arXiv Generative Models for Video G $0\le m(z)\le 1$ 是时空mask,选择背景$b$和前景$f$ $\odot$是 element-wise multiplication 训练时,加入$\lambda||m(z)||_ 1$来使$m$稀疏,也就是尽量用$b$少用$f$ D solve two problems classify realistic scenes from synthetically generated scenes recognize realistic motion between frames