This document shows and explains the process for compressing images, in grayscale and in color, using two neural topologies: image function and funnel. For the analysis of neuronal schemes, the number of neurons and layers, type of image, size and number of blocks during training are considered; in order to give experimental support to neural architectures. Quality criteria of the image obtained are also analyzed, such as peak signal-to-noise ratio (PSNR) and compression rate. The importance of the selection of parameters evaluated in quality and compression time is evident. The experimentation process shows that the funnel-type architecture allows to achieve values ​​higher than 35dB in terms of PSNR and 2 bits per pixel in gray images or 3 bpp in color images, with times less than 3 seconds for images smaller than 1 mega pixel. Finally, some recommendations are made based on the methodology used when it is desired to understand images with feed-through networks around the selection of architecture parameters, the pre-processing of the image and the training of the network.