Random field models are able to synthesize a large variety of complex patterns with a small number of parameters. This paper discusses the use of a Gibbs random field model as part of an image coding system. In particular, some semantic and perceptual attributes of this model are addressed.