Abstract Visual pattern memory of Drosophila melanogaster at the torque meter is investigated by a new learning paradigm called novelty choice. In this procedure the fly is first exposed to four identical patterns presented at the wall of the cylinder surrounding it. In the test it has the choice between two pairs of patterns, a new one and one the same as the training pattern. Flies show a lasting preference for the new figure. Figures presented during training are not recognized as familiar in the test, if displayed (i) at a different height, (ii) at a different size, (iii) rotated or (iv) after contrast reversal. No special invariance mechanisms are found. A pixel-by-pixel matching process is sufficient to explain the observed data. M inor transfer effects can be explained if a graded similarity function is assumed. Recognition depends upon the overlap between the stored template and the actual image. The similarity function is best described by the ratio of the area of overlap to the area of the actual image. The similarity function is independent of the geometrical properties of the employed figures. Visual pattern memory at this basic level does not require the analysis of shape.