I will present recent progress in the extraction of transverse-momentum dependent
parton distribution functions (TMDs) with a particular emphasis on the role of neural network (NN) parametrizations and lattice QCD inputs. First, I will discuss the first proof of-concept extraction of unpolarised TMDs at next-to-next-to-next-to-leading logarithmic N3LL) accuracy using neural networks. By providing a flexible and data-driven framework, neural networks overcome several of the limitations inherent in traditional functional forms and yield an improved description of Drell-Yan data. This establishes the feasibility of NN-based TMD extractions and motivates further developments in this direction. I will then turn to the first joint study of the Collins-Soper kernel combining inputs from lattice QCD and TMD phenomenology. Using recent continuum-extrapolated lattice calculations at multiple lattice spacings, we assess their impact on a phenomenological extraction based on neural network parametrizations. Both Bayesian reweighting and, for the first time, a direct global fit including the 21 lattice points together with about 500 experimental measurements are performed. The inclusion of lattice data has an impact both on the central value and on the uncertainties of the extracted Collins-Soper kernel.