The inverse problem of three-dimensional (3D) low-frequency electromagnetic (EM) data is usually formulated as unconditional minimization of the Tikhonov parametric functional. The Gauss-Newton method ensures fast convergence, but has high computational and/or memory complexity due to the need to factorize the Hessian matrix. This difficulty can be overcome only partially by the use of modern massively parallel distributed memory clusters (for example, Grayver et al., 2013; Wang et al, 2018). The nonlinear conjugate-gradient (NLCG) or L-BFGS methods are less demanding in terms of computational load and memory consumption, but may suffer from slow convergence at complicated models. There are two approaches, in which a special kind of transformation of model parameters is proposed: a diagonal preconditioner of Newman and Boggs (2004) and the integral-sensitivity approach of Zhdanov (2002). These approaches are essential to achieve a tolerable convergence rate and, in fact, are very similar. Recent examples with these methods include (Commer and Newman, 2008) and (Čuma et al., 2017).