This work proposes a novel variation to source-free domain adaptation (SFDA) that achieves generalizability through interdomain-intraclass manifold fusion and encoding-guided clustering for classifying new and unseen categories. Multidomain adaptation is accomplished without accessing the source data using two stages; (i) interdomain-intraclass manifold fusion (IMF) and (ii) interclass-cluster-cohesive fine-grained classification (IFC). IMF stage adheres to real-world source data privacy concerns related to proprietary and intellectual property rights. In the IMF stage, intermediate embeddings generated from the source model (without accessing source domain data) are used to propagate source domain knowledge. Instead of linearly combining different classes, which can diminish class discriminability and is non-intuitive in the real world, we introduce interdomain-intraclass embedding mixup to filter the domain invariant class-specific features. The IFC stage enforces strong intraclass cohesion and interclass separation. We conduct our experimental analysis on the fine-grained vehicle detection (FGVD) dataset, a complex and chaotic unconstrained road dataset.