Here is an experiment inspired by this conv.js demo. It is a simple model to predict the RGB color value of a given image from x, y coordinate as in the following figure.
Once the model got trained, we can generate new images by feeding XY values. It’s nothing interesting to replicate the original image using the model, right? The point is that we can intentionally ignore(nullify) some of the layers in the trained model so that the model starts generating abstract images loosely based on the pattern in the original image.