What are the system requirements for Stable Diffusion?
Stable Diffusion runs best on an NVIDIA GPU with at least 6-8GB VRAM (12GB+ recommended for SDXL/SD 3.5), 16GB system RAM, and about 10-30GB free disk space. It works on Windows, Linux, and Apple Silicon Macs. AMD GPUs and CPU-only setups work but are slower. VRAM is the single most important spec.
Why — the first-principles explanation
The one spec that decides everything is VRAM, the memory on your graphics card. During generation the entire model plus the working image must fit in VRAM. If it fits, generation is fast; if it doesn't, the software either fails or falls back to slow tricks. That's why VRAM matters more than raw speed.
Different models have different appetites. SD 1.5 is lean and runs on 4-6GB. SDXL roughly doubles the size and wants 8-12GB. The newest SD 3.5 Large is bigger still and prefers 12-24GB, though smaller variants exist. So your VRAM sets a ceiling on which models you can comfortably run.
The GPU brand matters too. NVIDIA is the default because its CUDA platform is what these tools were built for, giving the widest support and best speed. AMD cards work through ROCm on Linux or DirectML on Windows but with more setup and lower performance. Apple Silicon uses Metal and runs well on M-series chips.
Everything else is secondary. System RAM of 16GB keeps things smooth, disk space matters because each model checkpoint is 2-7GB and they add up fast, and a modern CPU helps loading. You can even run CPU-only with no GPU, but a single image may take several minutes instead of seconds. In short: a mid-range NVIDIA card with 8GB+ VRAM is the sweet spot.
An example that makes it click
Think of VRAM as the size of your kitchen counter. The model and the image you're cooking both need to sit on that counter at once. A tiny counter, 4GB, fits a simple recipe, SD 1.5, but not a big feast. A large counter, 12GB, lets you spread out a complex dish like SDXL without knocking things off.
The stove's brand is like NVIDIA versus AMD: they both cook, but recipes were written for the NVIDIA stove, so it just works, while the AMD stove needs a few adjustments first.
How to do it
- Check your GPU's VRAM: aim for 8GB+ for SDXL, 12GB+ for SD 3.5, 4-6GB minimum for SD 1.5.
- Prefer an NVIDIA GPU for the best compatibility and speed via CUDA.
- Ensure at least 16GB of system RAM.
- Free up 10-30GB of disk space for the interface and model checkpoints (each 2-7GB).
- On AMD, plan for ROCm (Linux) or DirectML (Windows); on Mac, use Apple Silicon with Metal.
- If you have no capable GPU, expect CPU generation to take minutes per image, or use a cloud/hosted option.
Key facts
- VRAM is the most important requirement: 4-6GB for SD 1.5, 8-12GB for SDXL, 12-24GB for SD 3.5 Large.
- NVIDIA GPUs offer the best support via CUDA; AMD needs ROCm (Linux) or DirectML (Windows).
- At least 16GB of system RAM is recommended for smooth operation.
- Model checkpoints are typically 2-7GB each, so 10-30GB free disk space is advisable.
- Runs on Windows, Linux, and Apple Silicon macOS; CPU-only works but takes minutes per image.
The open-source image model you can run on your own hardware.
Affiliate link — we may earn a commission at no cost to you.▶ The 60-second explainer (script)
What do you need to run Stable Diffusion? One spec matters more than all the others: VRAM, the memory on your graphics card. The whole model has to fit in it. Here's the rule of thumb. For the lightweight SD 1.5, four to six gigabytes of VRAM is enough. For the sharper SDXL, aim for eight to twelve. For the newest SD 3.5, twelve to twenty-four is ideal. NVIDIA cards are the safest choice, because these tools were built for NVIDIA's CUDA platform, so they just work. AMD cards run too, but need extra setup, and Apple Silicon Macs work great through Metal. Beyond the GPU, get sixteen gigabytes of system RAM, and keep ten to thirty gigabytes of free disk space, because each model file is several gigabytes and they pile up. No powerful graphics card? You can still run it on your CPU alone, but a single image might take several minutes instead of seconds. The sweet spot for most people is a mid-range NVIDIA GPU with eight gigabytes or more.
What authoritative sources say
People also ask
What's the minimum GPU for Stable Diffusion?
An NVIDIA GPU with about 4-6GB VRAM can run SD 1.5 with optimizations. For SDXL, aim for 8GB or more.
Do I need an NVIDIA card?
It's not required but strongly recommended. NVIDIA has the best support and speed. AMD and Apple Silicon work with extra setup.
How much disk space does it use?
Plan for 10-30GB. The interface is small, but each model checkpoint is 2-7GB and they accumulate quickly.
Can I run it without a graphics card?
Yes, on CPU only, but it's slow, often several minutes per image. A GPU or a hosted service is far more practical.