So I finally caved and started watching those videos about people making these insane AI images and I am literally obsessed. I have been playing around with some free stuff online but it takes forever and honestly the results are not even that great. I really want to build my own pc setup so I can run stuff locally and stop waiting in queues or whatever. I have zero idea what I am doing though so please be patient with me.
I have about 800 to 1000 bucks saved up for the card alone because I heard that the graphics card is basically the engine for this stuff right? I live in a tiny apartment in Chicago so I dont need some massive server rack thing but I want something that wont choke when I try to generate portraits or maybe some animation stuff later on. Is NVIDIA still the way to go? People keep saying something about VRAM and honestly that sounds like a foreign language to me. Does more memory actually matter more than the speed or what?
Im kinda hoping to get this sorted by next month because I have some time off work and I just want to sit in my room and generate weird art. I looked at the RTX 4070 or maybe the 4080 but I dont even know if those are overkill or if I am looking at the wrong thing entirely. Are there specific brands I should avoid or is a card a card? This whole thing is super confusing but I am so excited to get started if I can just figure out what to plug into my motherboard. What is the absolute best card I can get for around a grand that wont make me want to cry when I try to run Stable Diffusion...
Welcome to the rabbit hole! Honestly, you heard right, VRAM is king for AI. Stable Diffusion loves memory because it needs to load those massive models, and if you run out of space, things just crawl or crash. Forget about raw speed for a second; if you have 8GB of VRAM you are gonna struggle with newer stuff pretty quick. Since you have about a grand, NVIDIA is definitely the move. Their CUDA cores are basically the industry standard for this stuff, and honestly, it just works without a headache. Here is what I would look at: