Following Henson and Penny's guide "ANOVAs and SPM,"
you should include the subject effects. (See section 4, "Two-way within-
subject ANOVAs," especially Figure 7.)
On the other hand, if you do it using "partitioned" rather than "pooled"
errors (see Henson and Penny), if I'm not mistaken for 2x2 you would get
three group-level models, each with only one contrast image per subject.
(Two main effects and one interaction effect.) Which would remove the
need to model subjects, and is more in accordance with standard
statistical practice. (Though at least going by posts to the mailing list
here, SPM users tend to use "pooled" error much more often.)