Yes, it's strongly influenced by clip-interrogator, but I revamped the algorithm quite a bit. I think it could be improved even further without resorting to fine-tuning the BLIP model.