In four experiments we examined the false recognition of faces within the context of a larger model of face recognition that accounts for the effects of typicality and distinctiveness. Using morphing image processing techniques, we created a set of distracter faces that were mathematical blends of two 'parent' faces. These were used to address possible blending mechanisms underlying the false recognition. To address temporal context effects in any blending process, during study the two parent faces were seen either sequentially or separated by at least 20 other faces. During test, the two parent faces were replaced by a morph distracter. We found very high false alarm rates to the morph distracters, but no effects of the temporal context manipulation. In a forced choice version, subjects were more likely to choose the distracter over the parent face if the two parents are similar to each other, which demonstrates the strength of the false recognition effect and is consistent with a blending mechanism. Recognition models based on Nosofsky's Generalized Context Model (GCM, Nosofsky, 1986) could account for some but not all aspects of the data. A new model, called SimSample, is developed based on elements from models in categorization and recognition. This model can account for the effects of typicality and distinctiveness, but still has difficulty accounting for the high false alarm rates to the morphs. A version that includes explicit prototype representations can account for the morphs, as long as the prototype strength is proportional to the similarity of the two parents. This mixed model is consistent with a blending mechanism that is more likely to occur between similar rather than dissimilar faces.