If I understood correctly they just added image preprocessing and then they feed that into llama.cpp...