TL;DR — Multimodal models look like a single system in the demo but are actually a pipeline of independently-trained encoders bolted onto a shared decoder. Each encoder carries its own training distribution, and in production the whole system is bottlenecked by whichever modality has the narrowe...
Source: [Dev.to](https://dev.to/aiexplore369zoho/multimodal-models-are-only-as-good-as-their-weakest-encoder-5274)