Perceptual learning is learning to see. For example in a bisection task, three parallel lines are presented with the central line slightly offset towards the right or the left outer line. Participants indicate the offset direction. Training gradually improves performance. Models of perceptual learning explain learning by synaptic changes determined by the learning algorithm and the stimulus presentation. In these models, learning cannot occur when the very same stimulus is presented in all training trials. Here we show that, surprisingly, humans can improve performance in such "impossible" conditions. We trained observers with a line bisection task where the central line was always exactly in the middle for all 4160 training trials. Participants were not told about the “zero offset” and were instructed to indicate the offset direction as in a normal bisection task. Surprisingly, performance improved with gains similar to "normal" bisection experiments where both the left and right offset are presented. These results cannot be explained by most of current models of perceptual learning and reproduce previous studies in the auditory domain (Amitay, Irwin & Moore 2006). We suggest that perceptual learning occurs by mental imagery in accordance with previous results (Tartaglia, Bamert, Mast & Herzog, 2009, 2012).