Visual stimuli are initially represented in a retinotopic reference frame. To maintain spatial accuracy of gaze (i.e. eye in space) despite intervening eye and head movements, the visual input could be combined with dynamic feedback about ongoing gaze shifts. Alternatively, target coordinates could be updated in advance by using the preprogrammed gaze-motor command ('predictive remapping'). So far, previous experiments have not dissociated these possibilities. Here we study whether the visuomotor system accounts for saccadic eye-head movements that occur during target presentation. In this case, the system has to deal with fast dynamic changes of the retinal input, and with highly variable changes in relative eye and head movements that cannot be preprogrammed by the gaze control system. We performed visual-visual double-step experiments in which a brief (50 ms) stimulus was presented during a saccadic eye-head gaze shift toward a previously flashed visual target. Our results show that gaze shifts remain accurate under these dynamic conditions, even for stimuli presented near saccade onset, and that eyes and head are driven in oculocentric and craniocentric coordinates, respectively. These results cannot be explained by a predictive remapping scheme. We propose that the visuomotor system adequately processes dynamic changes in visual input that result from self-initiated gaze shifts, to construct a stable representation of visual targets in an absolute, supraretinal (e.g. world) reference frame. Predictive remapping may subserve transsaccadic integration, thus enabling perception of a stable visual scene despite eye movements, while dynamic feedback ensures accurate actions (e.g. eye-head orienting) to a selected goal.