Stable Diffusion 1.5 VRAM Requirements Guide
The lightweight classic. Runs on almost any modern 4GB–6GB GPU.
>98.2% retention. 0.9 GB weights, leaving +8.4 GB headroom on Intel Arc B580 12GB.
Near-lossless 8-bit precision requiring 4.2 GB VRAM for benchmark evaluations.
Uncompressed FP16 weights taking 5.3 GB. Demands multi-GPU or datacenter accelerators.
Architectural Memory Breakdown & Attention Scaling
Stable Diffusion 1.5 is the classic 1.0B parameter latent diffusion model built on an 860M parameter UNet paired with a CLIP ViT-L/14 text encoder. Operating natively at 512x512 resolution, its uncompressed FP16 weights consume just 2.1 GB, with peak sampling memory reaching ~3.4 GB. It runs smoothly on budget 4GB–6GB consumer GPUs and older graphics cards.
Enable PyTorch 2.0 scaled dot-product attention (`--opt-sdp-attention`) in your web UI. This eliminates peak attention memory spikes, allowing 512x512 image generation inside 3.2 GB of VRAM at over 15 iterations per second.
Intel Arc B580 12GB provides 12.0 GB physical VRAM (+8.4 GB operational margin at 8k context).
Run Stable Diffusion 1.5 Locally
ollama run sd-1-5:q4_K_Mvllm serve Runway / Stability AI/sd-1-5 --max-model-len 8192 --gpu-memory-utilization 0.95Frequently Asked Questions
What is the minimum GPU needed to run Stable Diffusion 1.5?
Any modern graphics card with at least 4GB of VRAM (e.g. GTX 1650, RTX 3050, or Apple Silicon Mac) can comfortably run SD 1.5 at 512x512 resolution.
Can SD 1.5 generate high-resolution 1024x1024 images?
While trained at 512x512, high-res fix or latent upscaling can generate 1024x1024 images, requiring ~5.5 GB peak VRAM during the upscaling pass.
Test Custom Hardware & Context Windows
Simulate Stable Diffusion 1.5 across all 31 GPUs, adjust context up to 77 tokens, or calculate RunPod break-even costs.
Open Stable Diffusion 1.5 in GPU Garden Matchmaker →