Optimization-based structure-from-motion (SfM) has been developed intensively for decades and has produced solid, well-understood results. Recently, end-to-end, feed-forward reconstruction pipelines have attracted intense interest, showing promising results on inputs where classical SfM tends to fail. This surge of interest makes it timely to revisit classical optimization-based methods and reconsider their role in the new era. In this talk, I will first revisit a global SfM method, GLOMAP, and then discuss how optimization and feed-forward models can be combined into a modern SfM pipeline, GLUEMAP, that draws on the strengths of both.