>>937550687 (OP)might be better suited for /g/ as that has people who are actually into technology and willing to help. /b/ is literally 99% israeli bots, haven't you seen the IP statistics during the shutdown?
to answer your question, JoyCaption is pretty good and lightweight, for local use. will accurately describe anything nsfw you throw at it. other than that, the keyword for "uncensored" LLMs isn't actually "uncensored" (that just means it was finetuned on nsfw materials, which isn't enough to bypass most internal safeguards), but "abliterated". any major open source LLM, whether with vision capabilities or not, has an abliterated version available.