[tempusalaria]

This is very similar to how LLMs are taught to understand images in llava style models (the image embeddings are encoded into the existing language token stream)