Researchers propose a method for retrieving historical images based on their content and the time period they were taken. The approach, called Temporally Decomposable Image Representations (TDIR), decomposes images into separate date and content components using orthogonal subspaces. This allows for transitive operations on embedding spaces, enabling the extraction of temporal information from one image to be injected into another without label supervision. The method is tested on a real-world problem of composed image retrieval on historical photographs, achieving competitive performance in both date estimation and object retrieval. --- Why it matters: This research matters because it provides a new way to organize and retrieve large collections of historical images based on their content and temporal context. This could be useful for applications such as photo archives, museums, or cultural heritage institutions. Source: https://arxiv.org/abs/2608.18694