>>105774449
we had this discussion a while ago. I went through a number of vision models via ollama/comfy but the results were pretty meh esp. with nsfw content, joy is alright but well its also pretty large. someone recommended I try gemini via api (semi-free with a generous daily quota) and it's amazing and zero censorship. also perfect for using it w/ large models because nothing needs to be loaded into the vram. (I haven't fed it real hc or other extreme content tho because I don't gen that shit). supposedly the large minicpm-v barely fits into 24gb vram and is pretty good.