In this paper we describe a strategy for efficient predictive compression of texture coordinates. Previous works in mesh compression often claim that this mesh property can simply be compressed with the same predictor that is already used for vertex positions. However, in the presence of discontinuities in the texture mapping such an approach results in unreasonable predictions. Our method avoids such predictions altogether. Rather than performing an unreasonable prediction, we switch to a less promising, but at least reasonable predictor. The resulting correctors are then compressed with different arithmetic contexts.