Full waveform inversion (FWI) is now well established as a standard tool for high resolution imaging in seismology and seismic exploration. Mathematically, it is formulated as a PDE-constrained non-linear non-convex inverse problem, where the discrepancy between observed data and synthetic data, solution of a wave propagation problem, is minimized. Large scale applications involve the use of local optimization methods, hence the risk to converge towards non-informative local minima. For this reason, decades of research have led to the proposition of various approaches to try to mitigate this issue and make full waveform inversion more robust. We might classify these approaches into 1) hierarchical approaches, where the data is interpreted progressively, piece by piece, 2) misfit function approaches, where the function used to compare the synthetic and observed data emancipates from the conventional leas-squares norm, 3) extension approaches, where the whole PDE-constrained formulation is modified in hope to generate a more convex inverse problem. In this presentation, we will see how we can connect all these approaches under the general concept of "extension" or "relaxation" of the FWI problem, and we will present results of such recent approaches on realistic case studies.