Stable Diffusion XL 1.0 (Base + Refiner) VRAM Requirements Guide
Native 1024x1024 UNet diffusion. Massive LoRA and Community ecosystem.
>98.2% retention. 2.2 GB weights, leaving +6.1 GB headroom on Intel Arc B580 12GB.
Near-lossless 8-bit precision requiring 7.6 GB VRAM for benchmark evaluations.
Uncompressed FP16 weights taking 10.8 GB. Demands multi-GPU or datacenter accelerators.
Architectural Memory Breakdown & Attention Scaling
Stability AI's Stable Diffusion XL (SDXL 1.0) is a 3.1B parameter latent diffusion model combining a 2.6B UNet core with dual CLIP text encoders (OpenCLIP ViT-bigG and CLIP ViT-L). Requiring 6.2 GB of FP16 weights and ~2.5 GB of peak VAE decoding memory during 1024x1024 synthesis, total peak allocation reaches ~8.8 GB, making 12GB GPUs the ideal sweet spot.
In Automatic1111 or ComfyUI, enable `--xformers` or PyTorch 2.0 SDPA attention and activate tiled VAE decoding (`--vae-tiling`). This reduces peak VAE memory from 2.5 GB down to 600 MB, preventing OOM errors on 8GB cards.
Intel Arc B580 12GB provides 12.0 GB physical VRAM (+6.1 GB operational margin at 8k context).
Run Stable Diffusion XL 1.0 (Base + Refiner) Locally
ollama run sdxl-1-0:q4_K_Mvllm serve Stability AI/sdxl-1-0 --max-model-len 8192 --gpu-memory-utilization 0.95Frequently Asked Questions
Can SDXL 1.0 run on an 8GB graphics card like the RTX 3060 Ti?
Yes! Enabling MedVRAM or PyTorch 2.0 SDPA with tiled VAE keeps peak memory below 7.5 GB, allowing reliable 1024x1024 image generation.
How much faster does SDXL run on 12GB–16GB GPUs compared to 8GB cards?
On 12GB+ GPUs, all model weights and VAE buffers remain resident in VRAM simultaneously, generating 1024x1024 images in 2–4 seconds without offloading.
Test Custom Hardware & Context Windows
Simulate Stable Diffusion XL 1.0 (Base + Refiner) across all 31 GPUs, adjust context up to 77 tokens, or calculate RunPod break-even costs.
Open Stable Diffusion XL 1.0 (Base + Refiner) in GPU Garden Matchmaker →