How much VRAM does SDXL need?

SDXL runs well on 8 GB of VRAM at 1024x1024, and is very comfortable on 12 GB. With memory-efficient attention you can push it onto 6 GB cards.

How much VRAM does FLUX.1 need?

FLUX.1 (12B) wants around 16-24 GB at full precision. With an 8-bit or GGUF-quantized build plus CPU offload it can run on 12 GB, and squeezed builds reach 8 GB at reduced speed.

Is more VRAM faster for image generation?

More VRAM lets you avoid offloading and generate at higher resolution or batch sizes, which is faster overall. But the GPU's raw speed matters too once the model already fits.

How Much VRAM for Stable Diffusion XL & FLUX? (2026 Guide)

Image models play by slightly different rules than LLMs: instead of a giant KV cache, the VRAM hit comes from the model itself plus the resolution you're generating at. The good news is SDXL is friendlier than its reputation. The bad news is FLUX earns its reputation.

The quick reference

Model	Comfortable VRAM	Minimum (with tricks)
Stable Diffusion 1.5	4–6 GB	Runs on almost anything
SDXL (6.6B)	8–12 GB	~6 GB with offload
SD 3.5 Large (8B)	12–16 GB	~8 GB quantized
FLUX.1 Schnell / Dev (12B)	16–24 GB	~8–12 GB quantized + offload
FLUX.2 Dev (32B)	24 GB+	Heavy — offload or quantize hard

Rule of thumb 8 GB = great SDXL machine. 12 GB = SDXL with headroom and quantized FLUX. 16–24 GB = FLUX at full quality without fighting the memory. Resolution and upscaling push every number higher.

Why FLUX is so much hungrier

FLUX.1 is a 12B-parameter transformer — roughly 8× the size of SD 1.5 — and ships at high precision. At full fidelity the model plus its text encoders can ask for 20 GB or more. The community has done heroic work shrinking it: 8-bit and GGUF-quantized FLUX builds, plus CPU offload of the text encoder, bring it down to 12 GB and even 8 GB cards, at the cost of speed and a little fidelity.

Resolution is the other dial

Generating at 1024×1024 is the baseline. Push to 1536² or run a hi-res upscale pass and VRAM climbs sharply — an upscale can add several gigabytes on its own. If you're on a tight card, generate at standard resolution and upscale as a separate, lighter step rather than in one giant pass.

Size a card for your image model

Pick SDXL, SD 3.5, or FLUX, choose your target resolution, and get the VRAM target plus a GPU that comfortably fits it.

Open the Local AI Calculator →

FAQ

Can I run FLUX on an 8 GB GPU?

Yes, with a quantized (GGUF/8-bit) build and CPU offload of the text encoder — but it's slow. A 12 GB card is a much happier minimum for FLUX.

Is SDXL or FLUX better for a 12 GB card?

SDXL runs natively and fast on 12 GB. FLUX is possible quantized but tighter. If speed matters, SDXL; if you want FLUX's prompt-following, accept the squeeze.

Does VRAM affect image quality?

Not directly — quality comes from the model and settings. But more VRAM lets you run the full-precision model and higher resolutions, which improves results.

We may partner with companies or groups to affiliate hardware products based on user needs, earning a commission from qualifying purchases. Image-model VRAM varies widely with pipeline, precision, offload settings, and resolution; figures are practical estimates. Data current as of June 2026.