The systematic study of morphings for non parametric shapes suffers from ambiguities in defining good general morphings, such as the trade-off between plausibility and smoothness, above all under large topology changes. In the recent years, only neural networks have offered a generic solution, using their latent space as a shape prior. But these models are optimized for single shape reconstruction, giving little control on the generated morphings. In this paper, we show how qualitatively similar results can be achieved when replacing neural networks with a set of carefully crafted components: a style-content separation method via the fitting of a Deformable Voxel Grid, a similarity metric adapted to the extracted content, and a formulation of morphings as minimal paths in a graph. While forgoing the automatic learning of a generative model, we still achieve similar morphing capabilities. We performed various evaluations, quantitative analysis on the robustness of our proposed method and on the quality of the results, and demonstrate the usefulness of each component. Finally, we provide guidance on how manual intervention can improve quality. This is indeed possible since, unlike neural networks, each component in our method is interpretable.