It's a lot easier to run Stable Diffusion locally. Meanwhile only the dumbest LLMs work on ordinary consumer GPUs. Datacenter GPUs with 80 GB vram are ridiculously expensive.