Stable Diffusion requires a minimum of 4GB of VRAM for SD 1.5, but modern models like Flux.1 Dev and SD 3.5 Large need 16-24GB for reliable generation. If your hardware falls short, a cloud GPU is the fastest path to running these models without buying new equipment.
This guide covers hardware requirements for every major Stable Diffusion model, how to pick the right model for your GPU, a step-by-step local install, and how to move your workflow to the cloud.
Stable Diffusion VRAM Requirements by Model
VRAM is the single most important factor in determining what you can run. If a model doesn't fit in VRAM, the workflow will spill into system RAM or fail entirely, causing out-of-memory errors or extremely slow generation.
| Model | Recommended VRAM | System RAM | Model Size on Disk | Notes |
|---|---|---|---|---|
| Stable Diffusion 1.5 | 6–8GB | 16GB | 2GB | Best for older hardware. Largest LoRA and ControlNet ecosystem. |
| Stable Diffusion XL 1.0 | 8–12GB | 16GB | 6.5GB | Higher resolution (1024x1024). More memory required for the Refiner and LoRAs. |
| Stable Diffusion 3.5 Large | 16GB (FP8) / 24GB (FP16) | 32GB | 16GB | Heavy and prone to memory errors on low-end cards without optimization flags. |
| VRAM figures include approximate overhead from text encoders and VAE. | ||||
How Much VRAM Do You Actually Need?
The practical answer depends on which models and features you plan to use. Here is what each VRAM tier unlocks.
4-8GB covers SD 1.5 at full quality and SDXL with memory optimization flags. This tier is sufficient for single-image generation without ControlNet or multiple LoRAs stacked simultaneously.
12-16GB is the comfortable floor for modern workflows. SDXL runs without workarounds and SD 3.5 Large becomes viable. This tier handles moderate LoRA stacks and single ControlNet passes.
24GB+ removes nearly every constraint for image generation. Complex ControlNet pipelines work without memory pressure, high-resolution upscaling passes can run without tiling, and training workflows (DreamBooth, LoRA fine-tuning) become feasible.
Stable Diffusion Minimum System Requirements
Beyond the GPU, several other components affect whether Stable Diffusion will run reliably.
- OS: Windows 11, macOS, or Linux
- GPU: NVIDIA (recommended for CUDA support); AMD and some Intel GPUs work with additional configuration
- VRAM: 4GB minimum for SD 1.5; 16GB practical floor for modern models
- System RAM: 16GB minimum; 32GB recommended when using CPU offloading with GGUF variants
- Storage: 20GB+ (model checkpoints alone range from 4GB to 41GB)
- Python: 3.10 or higher
- CUDA: 12.x recommended for NVIDIA GPUs
How to Install Stable Diffusion Locally
Local installation takes under 30 minutes once you have Python and your GPU drivers in place. The steps below use ComfyUI, which is the most capable and actively maintained frontend in 2026.
Step 1: Choose a Frontend
Three frontends dominate the Stable Diffusion ecosystem. The right one depends on how much control you want over the generation pipeline.
| Frontend | Interface | Best For | VRAM Overhead |
|---|---|---|---|
| ComfyUI | Node graph | Complex workflows, video, reproducibility | Low |
| Forge Neo | Tab-based | Beginners, SDXL, lower VRAM usage than A1111 | Low |
| AUTOMATIC1111 | Tab-based | Familiar UI, large extension ecosystem | Medium |
ComfyUI is the recommended starting point in 2026. Its node graph exposes every step of the pipeline and makes complex workflows shareable as JSON files. Forge Neo is the right alternative if you prefer a traditional web UI with lower memory usage than AUTOMATIC1111.
Step 2: Install ComfyUI
Download the ComfyUI desktop installer for Windows, or clone the repository and install Python dependencies manually on macOS and Linux. You will need Python 3.10 or higher and a CUDA-compatible GPU driver before running the installer.
Once installed, open ComfyUI Manager from the interface and click Update ComfyUI to ensure you are on the latest build. ComfyUI Manager also handles all custom node installations, so install it before downloading any additional packages.
Step 3: Download a Model Checkpoint
Model weights are hosted on Hugging Face and CivitAI. Download the .safetensors file for your chosen model and place it in the correct subdirectory:
- Standard checkpoints (SD 1.5, SDXL):
ComfyUI/models/checkpoints/ - Flux and Z-Image diffusion models:
ComfyUI/models/diffusion_models/ - Text encoders (T5, CLIP):
ComfyUI/models/text_encoders/ - VAE files:
ComfyUI/models/vae/
After placing the files, restart ComfyUI. The model will appear in the relevant loader node dropdown.
Step 4: Load a Workflow and Generate
ComfyUI opens to a default text-to-image workflow. Click Browse Example Workflows in the left sidebar to find official templates for SD 1.5, SDXL, Flux, and more.
Enter your prompt, set the image dimensions, and click Queue Prompt. Generation time depends on your GPU and the number of inference steps configured.
Fixing CUDA Out-of-Memory (OOM) Errors in Stable Diffusion
CUDA out-of-memory (OOM) errors are the most common problem new users encounter. They happen when the model and pipeline cannot fit within your GPU's VRAM.
- Reduce output resolution. Moving from 1024x1024 to 768x768 cuts VRAM usage significantly. If OOM errors persist, try 512x512 as a baseline.
- Launch with
--lowvram. Enables sequential component processing at a 20-30% speed penalty. Use this if resolution reduction alone doesn't resolve the error. - Switch to FP8 checkpoints. Reduces VRAM usage by roughly 40-50% with minimal quality loss. Both ComfyUI and Forge Neo support FP8 natively.
- Use GGUF-quantized models. The most aggressive reduction for 6-8GB cards, bringing larger models within reach at a moderate quality cost.
If OOM errors persist, a cloud GPU removes the hardware constraint entirely. See the Thunder Compute ComfyUI cloud setup guide.
How to Run Stable Diffusion in the Cloud
Cloud GPUs let you run any model at full precision without buying new hardware. They are the practical option when your local GPU doesn't have enough VRAM, or when you're running large batch jobs.
Thunder Compute provides GPU instances optimized for generative AI workflows, starting at $0.35/hr for an RTX A6000 with 48GB of VRAM. A pre-configured ComfyUI template is available that eliminates the local installation steps described above: no Python environment setup, and no driver configuration.
The Quick Runthrough:
Getting started is simple. Here is a brief overview:
- Select Your Instance: Log into Thunder Compute and choose a GPU.
- Deploy the Template: Choose the ComfyUI Template. This pre-installs necessary drivers.
- Launch & Create: Open the provided link to your workspace and start generating immediately.
Our ComfyUI setup guide offers a detailed walkthrough on setting up your environment, or try Forge Neo for a simpler web UI alternative.
GPU Options for Stable Diffusion on Thunder Compute
| GPU | VRAM | Price | Best Use Case |
|---|---|---|---|
| RTX A6000 | 48GB | $0.35/hr | SDXL |
| NVIDIA A100 | 80GB | $1.09/hr | LoRA training |
| NVIDIA H100 | 80GB | $2.19/hr | High-volume batch generation, large-scale training runs |
Per-minute billing means you only pay for active generation time. For bursty workflows like prompt testing and model iteration, this typically saves 30-40% compared to hourly billing increments.
Last Thoughts on Running Stable Diffusion
SD 1.5 still runs on 4GB of VRAM, but SDXL, Flux.1 Dev, and the newer 2026 models deliver better results and demand more memory. If your local hardware falls short, cloud GPUs are the fastest path to running any model at full quality. Thunder Compute's RTX A6000 instances cover every current image model at $0.35/hr, with a one-click ComfyUI template that skips local setup.
Start generating on Thunder Compute GPUs today. No hardware required.