Mullerian mimicry is a widely studied phenomenon in evolutionary biology. The theory predicts that when multiple warning signals coexist, they will tend to converge to the most common signal through frequency-dependent selection. Heliconius butterflies are a renowned example of Mullerian mimicry yet they present a peculiar paradox, in that several Heliconius species have diverged into an array of distinct warning phenotypes, instead of converging on any one specific warning coloration. The result is that in any specific location numerous unique mimicry rings can be found. However, some species have evolved the ability to maintain multiple mimetic warning phenotypes in a single population, a phenomenon known as polymorphic mimicry. In these populations, the distinct morphs actually belong to different local mimicry rings. The selective pressures that allow polymorphic mimicry to evolve and be maintained remain largely unresolved. Using artificial models of Heliconius doris, which displays such polymorphic mimicry, we quantify predation across sites with varying mimicry rings in French Guiana and Panama. We use these predation experiments to test specific predictions about the selective pressures involved in the maintenance of polymorphic mimicry in H. doris. We find clear evidence of frequency-dependent selection maintaining alternative morphs in the presence of differing mimicry rings. We also find evidence of divergent selection on warning color patterns, resulting in marked difference of morphs in different geographic populations. We propose that these predation forms not only contribute to the maintence of polymorphic mimicry within H. doris, but also may promote intraspecific warning diversity of co-mimics, thereby providing insight into the selective forces driving the paradox of warning color diversity among Heliconius butterflies.