Conditional coding is one of the many opportunities, neural networks bring to the field of image and video compression. It is applicable whenever some information about the current frame is known at both encoder and decoder. Examples of such information can be the spatial information obtained from previously transmitted image areas, temporal information from previously transmitted frames, or structural information which was transmitted over a side-channel. In this thesis, we examine neural-network-based methods to exploit those three kinds of information. We cover a conditional coding method which functions as an intra prediction system, effectively achieving multi-mode prediction using only a single prediction network. Furthermore, we show that conditional inter-frame coding outperforms the traditional residual coding method. Finally, we design and analyze a multi-scale autoencoder which can exploit the knowledge of structural information, which can use rate-distortion optimization to achieve coding gains.weiterlesen