How much VRAM do I need for SDXL?

SDXL needs roughly 7GB of VRAM for the model plus activations at 1024px, so an 8GB card runs it but with little headroom, and 12GB is comfortable. Adding a ControlNet pushes the comfortable minimum toward 12GB or more.

Can I run Flux.1 on a consumer GPU?

Flux.1 in fp16 wants around 22GB, which exceeds most consumer cards, but the fp8 build drops to roughly 12GB and runs on a 16GB GPU comfortably or a 12GB card tightly. Quantization is the key to fitting Flux on consumer hardware.

Why does resolution change the requirement so much?

Activation memory scales with pixel area, so doubling each dimension quadruples the activations even though the model weights stay the same. Generating at 1536px instead of 1024px is a large VRAM jump for the same model.

What if the calculator says it won't fit?

Lower the resolution or batch size, switch to an fp8 or quantized checkpoint, or enable offloading (--medvram / --lowvram) and tiled VAE. These reduce peak usage enough to fit most models on smaller cards at the cost of speed.

GPU VRAM Requirements Calculator

GPU VRAM requirements calculator

The first question with any local image model is blunt: will it even run on my card? VRAM is the hard limit — run out and generation crashes outright. The answer depends on the model, the resolution, the batch size, and whether you stack a ControlNet on top. This calculator estimates the peak VRAM a configuration needs and checks it against your GPU so you know before you download an 8GB checkpoint.

How it works

Each model has a baseline weight footprint at its native resolution. The calculator scales the activation memory by pixel area — doubling resolution quadruples activations — and by batch size, since each extra image adds its own latent. A ControlNet loads a second model and adds a fixed chunk on top. The total is compared to your card’s VRAM, and the verdict tells you whether you have headroom, are running tight, or fall short and by how much.

Tips when you’re short on VRAM

Drop resolution first. It’s the biggest lever — 768px instead of 1024px cuts activations by nearly half on the same model.
Use fp8 / quantized weights. Flux fp8 nearly halves the requirement versus fp16 with minimal quality loss.
Enable offloading. --medvram and --lowvram shuttle weights to system RAM; you trade speed for the ability to run at all.
Shrink the batch. Generating one image at a time frees the most memory — queue a batch sequentially instead of in parallel when VRAM is tight.