The uncanny valley (UV) hypothesis suggests that increasingly human-like robots or virtual characters elicit more familiarity in their observers (positive affinity) with the exception of near-human characters that elicit strong feelings of eeriness (negative affinity). We studied this hypothesis in three experiments with carefully matched images of virtual faces varying from artificial to realistic. We investigated both painted and computer-generated (CG) faces to tap a broad range of human-likeness and to test whether CG faces would be particularly sensitive to the UV effect. Overall, we observed a linear relationship with a slight upward curvature between human-likeness and affinity. In other words, less realistic faces triggered greater eeriness in an accelerating manner. We also observed a weak UV effect for CG faces; however, least human-like faces elicited much more negative affinity in comparison. We conclude that although CG faces elicit a weak UV effect, this effect is not fully analogous to the original UV hypothesis. Instead, the subjective evaluation curve for face images resembles an uncanny slope more than a UV. Based on our results, we also argue that subjective affinity should be contrasted against subjective rather than objective measures of human-likeness when testing UV.