In each of three visual pattern masking experiments, four curved letters (C, O, Q, S) and four angular letters (E, I, L, T) served as targets preceded or followed by either a curved mask (Q,S, and C superimposed) or an angular mask (T and E superimposed). With a dark fixation and interstimulus interval field and target-mask engergies that produce clearly identifiable targets, the following figural relationship effects were found. At stimulus onset asynchronies (SOAs) from 0 to 20 msec (in both forward and backward masking), target recognition was more accurate when targets and masks overlapped exactly (same features) than when they did not (different features). At backward masking SOAs beyond 20 msec, this pattern was reversed, but there was no such reversal in forward masking. Such results indicate that the dominant mechanism of masking at SOAs from 0 to 20 msec is luminance summation over time but that luminance summation gives way to feature-specific interference at longer SOAs. Subsequent experiments demonstrate that (a) luminance summation effects are reduced by using bright fixation and interstimulus fields and (b) feature-specific interference is eliminated by using low-energy (and, therefore, less than perfectly identifiable) targets and masks.