SDXL VRAM requirements (Stable Diffusion XL base 1.0)

Short answer: SDXL base 1.0's FP16 files total 6.94 GB (the UNet is 5.14 GB); the FP32 files are 13.88 GB.^[1] Stability AI publishes no official VRAM figure. Its model card says that if you're limited by GPU VRAM, call pipe.enable_model_cpu_offload().^[2]

SDXL base 1.0 is Stability AI's 2023 text-to-image model, licensed CreativeML Open RAIL++-M. Hugging Face lists 2,567,463,684 parameters for its UNet, and the repo was created on 2023-07-25.^[1] An optional refiner model is a separate download.^[2]

Weights only: file sizes on Hugging Face

From the stabilityai/stable-diffusion-xl-base-1.0 file listing (1 GB = 1,000,000,000 bytes; checked November 4, 2026).^[1]

The single-file sd_xl_base_1.0.safetensors used by ComfyUI and other UIs is also 6.94 GB.^[1]

Memory while generating

Activation memory comes on top of the weights and grows with resolution and batch size (the model is trained for 1024×1024 images). Stability publishes no measured figure at a named setup, and we haven't measured one, so we don't give a number. The official ways to cut memory on the model card are model CPU offload and, on older PyTorch, xformers memory-efficient attention.^[2]

What this means for your GPU

• The FP16 weights (6.94 GB) fit within the VRAM of 8 GB cards and up, with room left for activations shrinking on smaller cards. See what you can run on 8 GB and 12 GB.
• Out of memory? Use CPU offload, or see the CUDA out of memory guide.
• Bigger image models: SD 3.5 · FLUX.1 schnell · Qwen-Image.

Frequently asked questions

How much VRAM does SDXL need?

The FP16 weights are 6.94 GB. Stability doesn't publish an official VRAM figure; generating at 1024×1024 needs extra memory for activations, and the card recommends CPU offload if you're short.

Is the SDXL file 6.9 GB or 13.9 GB?

6.94 GB in FP16 (what most people download) and 13.88 GB in FP32.

Every number on this page links to its source.