Recommender systems (RS) traditionally leverage the large amount of user–item interaction data. This exposes RS to a lower recommendation quality in cold-start scenarios, as well as to a low recommendation quality in terms of beyond-accuracy evaluation metrics. State-of-the-art (SotA) models for cold-start scenarios rely on the use of side information on the items or the users, therefore relating recommendation to multimodal machine learning (ML). However, the most recent techniques from multimodal ML are often not applied to the domain of recommendation. Additionally, the evaluation of SotA multimodal RS often neglects beyond-accuracy aspects of recommendation. In this work, we outline research into designing novel multimodal RS based on SotA multimodal ML architectures for cold-start recommendation, and their evaluation and benchmark with preexisting multimodal RS in terms of accuracy and beyond-accuracy aspects of recommendation quality.