This paper presents HF-Net, a hierarchical localization approach that combines learned features to achieve robust and accurate 6-DoF localization in large-scale environments. The method leverages a coarse-to-fine paradigm, first performing a global retrieval to obtain location hypotheses and then matching local features within those candidate places. This approach significantly reduces runtime and makes the system suitable for real-time operation. HF-Net is a monolithic CNN that jointly predicts local features and global descriptors, maximizing computation sharing. The authors also introduce multitask distillation to train the network flexibly using multiple state-of-the-art predictors, achieving fast, robust, and accurate localization. The method sets a new state-of-the-art on challenging benchmarks for large-scale localization, demonstrating superior performance in various conditions, including day-night queries and substantial appearance variations.This paper presents HF-Net, a hierarchical localization approach that combines learned features to achieve robust and accurate 6-DoF localization in large-scale environments. The method leverages a coarse-to-fine paradigm, first performing a global retrieval to obtain location hypotheses and then matching local features within those candidate places. This approach significantly reduces runtime and makes the system suitable for real-time operation. HF-Net is a monolithic CNN that jointly predicts local features and global descriptors, maximizing computation sharing. The authors also introduce multitask distillation to train the network flexibly using multiple state-of-the-art predictors, achieving fast, robust, and accurate localization. The method sets a new state-of-the-art on challenging benchmarks for large-scale localization, demonstrating superior performance in various conditions, including day-night queries and substantial appearance variations.