There are two classes of reasons that can lead to a divergence from realism: efficiency reasons, and aesthetic reasons. Taking into account perception has large potentials to produce not only more realistic pictures, but also clearer, more relevant or more beautiful pictures

A large class of pictures need to represent reality in a schematic way to improve clarity. Maps, manuals or scientific illustration are examples where a non-photorealistic picture is often clearer than a photograph, because it can emphasize the relevant information and provide straighter clues. Understanding how schematic visual information is perceived and how visual clues are processed are crucial issues to produce more efficient illustrations. Similarly, a major bottleneck of 3D authoring is the difficulty to deal with a 3-dimensional space and the resulting user-interface problems. Effective solution implies to produce interactive images that provide the human perceptual system with functionally optimal inputs.

Since the landmark writings of Von Helmholtz, Arnheim and Gombrich, studying the human visual system has proved fruitful to understand issues and techniques of the visual arts. Our perception of 3D scenes explains the limitations of linear perspective and why alternative projections can be desirable. The different levels of cognitive representation of a scene can explain dramatically different styles of depiction. Gestalt psychology provides fruitful insights to the art of image composition. Low-level and high-level cognitive mechanisms help us understand how an image can attract the gaze of the beholder to important regions. Other examples are numerous.

Most of these results are unfortunately mainly qualitative and will be challenging to translate into automatic image-generation techniques. Truly beautiful images will remain the unutterable domain of sensitive artists. Nonetheless, the insight into art techniques offered by perception can help improve the quality of images produced by non-artists. This includes offering more relevant high-level controls that directly relate to the aesthetic dimension of images. Another potential avenue is to use simple computational models of perception to provide the user with crude feedback and alarms about the quality of images, e.g. with respect to balance or gaze attraction. The development of such techniques has great cross-fertilization potentials between perception and computer graphics, and can help to devise and validate more quantitative models.