Exploiting spatial redundancy in images is responsible for a large gain in the performance of image and video compression. The main tool to achieve this is called intra-frame prediction. In most state-of-the-art video coders, intra prediction is applied in a block-wise fashion. Up to now angular prediction was dominant, providing a low-complexity method covering a large variety of content. With deep learning, however, it is possible to create prediction methods covering a wider range of content, being able to predict structures which traditional modes can not predict accurately. Using the conditional autoencoder structure, we are able to train a single artificial neural network which is able to perform multi-mode prediction. In this paper, we derive the approach from the general formulation of the intra-prediction problem and introduce two extensions for spatial mode prediction and for chroma prediction support. Moreover, we propose a novel latent-space-based cross component prediction. We show the power of our prediction scheme with visual examples and report average gains of 1.13% in Bjontegaard delta rate in the luma component and 1.21% in the chroma component compared to VTM using only traditional modes.