Single-particle cryo-electron microscopy (cryo-EM) has emerged as a transformative technique for determining the three-dimensional structures of macromolecular complexes at near-atomic resolution. Its ability to visualize biomolecules in multiple functional states without the need for crystallization has provided unprecedented insights into their structure, dynamics, and mechanisms, making it a cornerstone in structural biology and drug discovery. Despite its success, cryo-EM faces several challenges that limit the achievable resolution and accuracy of reconstructions. Chief among these are the inherently low signal-to-noise ratio (SNR) of raw micrographs, the difficulty in accurately estimating particle orientations (pose estimation), and the presence of conformational and compositional heterogeneity in the sample. In recent years, deep learning has emerged as a leading approach for addressing these limitations, offering powerful methods for denoising, pose refinement, and disentangling structural variability. In this work, a method designed to exploit particle heterogeneity for iterative pose refinement is presented. The approach integrates two state-of-the-art tools : cryoDRGN, which models structural variability using deep generative networks, and Frealign, which performs high-resolution 3D refinement. These tools were combined into a unified pipeline and tested on real cryo-EM datasets, demonstrating the potential of the method to improve both the accuracy of pose estimation and the quality of heterogeneous reconstructions.