For LLMs, I’ve had really good results running Llama 3 in the Open Web UI docker container on a Nvidia Titan X (12GB VRAM).

For image generation tho, I agree more VRAM is better, but the algorithms still struggle with large image dimensions, ao you wind up needing to start small and iterarively upscale, which afaik works ok on weaker GPUs, but will gake problems.