Hyperbolic space has emerged as a promising alternative to Euclidean space for embedding high-dimensional data, including images. In particular, Hyperbolic embeddings have shown to be more effective in discovering hierarchical relationships between data points. However, experiments performed thus far have provided the model with explicit access to this hierarchy during training. This work shows Poincaré embeddings’ ability to discover class-subclass relationships in image datasets without direct supervision. In addition, we present applications of this result to content-based image retrieval, demonstrating that Poincaré embeddings can achieve better performance when the relevance of the retrieved elements is measured by taking the class hierarchy into account. Furthermore, we found that Hyperbolic embeddings outperform their Euclidean counterpart for complex class hierarchies. Finally, we discuss the limitations of these results and outline future research directions.