Shouldn't be too difficult, but it would be hard to get it going in real time on affordable local hardware without using heavy quantization. 

Set up a venv and run insightface on it's own backend that also hosts your browser front end. Have the backend grab a webcam cap. Next, call it in to a open ai endpoint and whatever multimodal model you fancy at the moment. Send along some cooked insightface data and the rest of your caption prompt.