Hugging Face has published a technical guide on training and finetuning multimodal embedding and reranker models with Sentence Transformers. The post is aimed at developers building retrieval systems that need to represent and rank data across multiple modalities.
Since there is only one source, there are no conflicting viewpoints to compare. The guide is notable because it treats embedding training and reranker finetuning as connected parts of the same pipeline, rather than as separate concerns.
The practical focus on Sentence Transformers suggests that the material is intended for practitioners who want to move from off-the-shelf models to custom-trained ones. The post likely walks through key steps for adapting models to specific multimodal retrieval tasks.