Current theories of visual imagery hold that the same neural processes govern both the representation of real objects and the representation of imagined (but real) objects. Here we test whether the representation of imagined (real) objects and the representation of imagined (but unreal) objects depend on the same or different neurocognitive processes. A likely clinical group for a dissociation between these two types of imagination are children with autism, since they show deficits in imaginative play, impoverished imagination is part of their diagnosis, but they can search for hidden objects. The present study explored imagination in autism using experimental methods. Experiment 1 investigated if children with autism could introduce changes to their representations of people and houses, using Karmiloff-Smith's (1989) technique of asking children to draw “impossible” people or houses. Results showed that children with autism were significantly worse than matched controls in their ability to introduce “unreal” changes to their representations of people and houses. Instead, they tended to draw real people or objects. Experiment 2 investigated whether the performance in Experiment 1 by children with autism was due to an inability to disengage from “real world” representations, as executive dysfunction theorists would argue. To do this, the experimenter instructed them on what to draw and how to draw it. Results showed that even when executive control passed to the experimenter in this way, the children with autism were still significantly impaired in their ability to draw imaginary but unreal things relative to the matched controls. Experiment 3 investigated whether the results from Experiments 1 and 2 arose because of a generativity deficit in autism, which might be the executive dysfunction theorists' alternative account. To test this, the same subjects were given a test of Verbal Fluency and a test of imagining multiple functions of a brick. Results showed that the children with autism were no worse than clinical controls in their ability to generate ideas about real objects, suggesting that a global generativity deficit cannot explain the previous findings. Rather, these results point to a specific impairment in the ability to imagine unreal objects. This is discussed in terms of its possible neural dissociability from other kinds of imagery, and in terms of its possible relationship to theory of mind.