Stable Diffusion needs at least 4GB of VRAM for SD 1.5, while newer models like Flux.1 Dev and SD 3.5 Large need 16-24GB for reliable generation. A cloud GPU is the fastest way to run these models when local hardware falls short.
This guide covers VRAM and system requirements for every major Stable Diffusion model, a model picker by GPU tier, a step-by-step local install, and cloud GPU options.
Takeaways
- VRAM decides what runs: SD 1.5 needs 4GB, SDXL needs 8GB, and SD 3.5 Large needs over 18GB at BF16.
- Official minimums exclude overhead: text encoders, LoRAs, and ControlNet push real usage above publisher figures.
- FP8 and GGUF quantization bring SD 3.5, Flux, and Z-Image Turbo within reach of 8-12GB cards.
- ComfyUI is the recommended frontend for running Stable Diffusion, Flux, and video models locally.
- Cloud GPUs remove the VRAM ceiling: Thunder Compute's RTX A6000 offers 48GB of VRAM at $0.35/hr.
Stable Diffusion VRAM Requirements by Model
VRAM is the single most important factor in what Stable Diffusion models you can run. A model that doesn't fit in VRAM spills into system RAM or fails entirely. The result is out-of-memory errors or much slower generation.
Publisher VRAM figures often understate real-world needs. Publishers state minimums for only some models, and those figures usually exclude text encoders, LoRAs, and ControlNet.
The VRAM table lists each publisher's figure next to:
- Practical minimums which assume FP8 (8-bit floating point) or GGUF (a compressed model format) quantization.
- Recommended figures that include approximate overhead from text encoders and the VAE.
| Model | Official VRAM | Practical Minimum | Recommended VRAM | System RAM | Model Size on Disk | Notes |
|---|---|---|---|---|---|---|
| Stable Diffusion 1.5 | Not published1 | 4GB | 6–8GB | 16GB | ~2GB | Best for older hardware. Large LoRA and ControlNet ecosystem. |
| Stable Diffusion XL 1.0 | 8GB2 | 6GB | 12GB (16GB with Refiner) | 16GB | ~7GB | Higher resolution (1024x1024). More memory required for the Refiner and LoRAs. |
| Stable Diffusion 3.5 Medium | 9.9GB, excluding text encoders3 | 8GB | 12–16GB | 16GB | ~5GB plus text encoders | Heavy and prone to memory errors on low-end cards without optimization flags. |
| Stable Diffusion 3.5 Large | 18GB4 | 12GB | 24GB | 32GB | ~16GB | Flagship SD 3.5 model |
| Stable Video Diffusion XT | Not published | 8GB (with --lowvram) |
16GB+ | 16GB | ~10GB | Image-to-video, not text-to-image. Generates 25 frames at 1024x576. |
| Newer Alternatives to Stable Diffusion | ||||||
| Flux.1 Dev | Not published | 12GB | 24GB | 32GB | ~24GB | Stronger prompt adherence than SD models. Runs on 12GB with FP8, or 8GB with GGUF at a quality cost. |
| Flux.2 Klein 4B | ~13GB5 | 8GB | 16GB | 16GB | ~8GB | Smallest Flux.2 model. Built for fast, interactive generation. |
| Z-Image Turbo | 16GB6 | 8GB | 16GB | 16GB (32GB for GGUF) | ~12GB plus ~7GB text encoder | Fast 8-step generation. GGUF variants fit 6GB cards but need more system RAM. |
| 1The 4GB figure comes from the AUTOMATIC1111 documentation. | ||||||
| 2 From Stability AI's SDXL launch specs: an NVIDIA RTX 20-series card or better with 8GB of VRAM and 16GB of system RAM. | ||||||
| 3 From Stability AI's Stable Diffusion 3.5 announcement. | ||||||
| 4 From NVIDIA and Stability AI's FP8 release, which cut VRAM use by 40%. | ||||||
| 5 From Black Forest Labs' Flux.2 Klein announcement. | ||||||
| 6 From the Z-Image Turbo model card, which states the model fits on 16GB consumer GPUs. | ||||||
Stable Diffusion XL (SDXL) Requirements
SDXL needs 8GB of VRAM at minimum, according to Stability AI's launch specs, plus 16GB of system RAM and an NVIDIA RTX 20-series card or newer. An 8GB card generates SDXL images at the model's native 1024x1024 resolution. However, 8GB leaves little headroom for LoRAs or ControlNet.
SDXL runs comfortably on 12GB. Loading the SDXL Refiner alongside the base model pushes usage toward 16GB, as do stacked ControlNets and LoRAs. FP8 checkpoints cut SDXL's footprint to roughly 6GB with minimal quality loss.
Stable Video Diffusion Requirements
Stable Video Diffusion (SVD) is a separate image-to-video model built on SD 2.1, not a mode of Stable Diffusion. SVD-XT generates 25 frames at 576x1024. SVD-XT needs 8GB of VRAM at minimum, and 16GB+ avoids out-of-memory errors from pipeline overhead.
Read nextStable Diffusion Video Generation VRAM requirements, ComfyUI setup, workflows and GPU needs.
Stable Diffusion Minimum vs Recommended System Requirements
The GPU matters most, but system RAM, storage, and software versions also determine whether Stable Diffusion runs reliably. The minimum specs run SD 1.5. The recommended specs run SD 3.5, SDXL, Flux, and multi-LoRA workflows without constant workarounds.
| Component | Minimum | Recommended |
|---|---|---|
| OS | Windows 10/11, macOS, or Linux | Windows 11 or Linux |
| GPU | NVIDIA GPU with CUDA support; AMD and some Intel GPUs work with additional configuration | NVIDIA RTX 30-series or newer |
| VRAM | 4GB | 12–16GB |
| System RAM | 16GB | 32GB (needed for CPU offloading with GGUF variants) |
| Storage | 20GB free | 100GB+ on an SSD (checkpoints range from about 2GB to 24GB+ each) |
| Python | 3.10 | 3.10 or higher |
| CUDA | 12.x-compatible NVIDIA driver | CUDA 12.x with the latest NVIDIA driver |
How to Install Stable Diffusion Locally
A local Stable Diffusion install takes under 30 minutes once Python and GPU drivers are in place. This guide's install steps use ComfyUI, one of the most flexible and actively maintained frontends in 2026.
Step 1: Choose a Frontend
Three frontends dominate the Stable Diffusion ecosystem. The right frontend 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. ComfyUI's node graph exposes every step of the pipeline and saves complex workflows as shareable JSON files. Forge Neo (a maintained fork of the Forge web UI) suits users who prefer a traditional web UI with lower memory usage than AUTOMATIC1111.
Step 2: Install ComfyUI
Download the ComfyUI desktop installer for Windows or macOS, or clone the repository and install Python dependencies manually on Linux. The desktop app sets up its own Python environment. Manual installs need Python 3.10 or higher and a CUDA-compatible GPU driver.
Open ComfyUI Manager after installation and click Update ComfyUI to get the latest build. ComfyUI Manager also handles custom node installations, so set it up before downloading 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 matching ComfyUI 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/
Restart ComfyUI after placing the files. The model then appears in the matching loader node dropdown.
Step 4: Load a Workflow and Generate
ComfyUI opens to a default text-to-image workflow. Open the Templates browser to find official workflows for SD 1.5, SDXL, Flux, and more.
Enter your prompt, set the image dimensions, and click Run. Generation time depends on your GPU and the number of inference steps.
Fixing CUDA Out-of-Memory (OOM) Errors in Stable Diffusion
CUDA out-of-memory (OOM) errors are among the most common problems for new Stable Diffusion users. OOM errors occur when the model and pipeline exceed your GPU's VRAM.
- Reduce output resolution. Dropping from 1024x1024 to 768x768 cuts VRAM usage significantly. Try 512x512 as a baseline if OOM errors persist.
- Launch with
--lowvram. The flag processes model components sequentially, which noticeably slows generation. Use--lowvramwhen lower resolution alone doesn't resolve the error. - Switch to FP8 checkpoints. FP8 cuts VRAM usage by roughly 40-50% with minimal quality loss. ComfyUI and Forge Neo both support FP8 natively.
- Use GGUF-quantized models. GGUF delivers the largest VRAM reduction for 6-8GB cards, bringing larger models within reach at a moderate quality cost.
Read nextA cloud GPU removes the VRAM constraint when OOM errors persist. See the Thunder Compute ComfyUI cloud setup guide.
How to Run Stable Diffusion in the Cloud
Cloud GPUs run large Stable Diffusion models at full precision without new hardware. Cloud GPUs suit users whose local GPU lacks VRAM and users running large batch jobs.
Thunder Compute offers GPU instances for generative AI workflows, starting at $0.35/hr for an RTX A6000 with 48GB of VRAM. A pre-configured ComfyUI template skips the local install steps entirely, with no Python environment or driver setup.
The Quick Runthrough:
Launching ComfyUI on Thunder Compute takes three steps:
- Select Your Instance: Log into Thunder Compute and choose a GPU.
- Deploy the Template: Choose the ComfyUI template, which pre-installs the necessary drivers.
- Launch and Create: Open the workspace link and start generating.
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 | $3.20/hr | High-volume batch generation, large-scale training runs |
Thunder Compute bills per minute, so you pay only for active generation time. This keeps costs down for bursty workflows like prompt testing and model iteration.
Last Thoughts on Stable Diffusion Requirements
Stable Diffusion requirements start at 4GB of VRAM for SD 1.5, but SDXL, Flux.1 Dev, and newer 2026 models deliver better results and need more memory.
Cloud GPUs are the fastest way to run larger models at full quality when local hardware falls short. Thunder Compute's RTX A6000 instances are a great starting point at $0.35/hr, with 48GB of VRAM and a one-click ComfyUI template.
Start generating on Thunder Compute GPUs today. No hardware required.
FAQ
What are the Stable Diffusion hardware requirements for 2026?
The minimum is an NVIDIA GPU with 4GB of VRAM running SD 1.5. Modern models need 12-16GB in practice. A100 or H100 cloud instances suit unquantized generation, LoRA training, and production batch workflows.
Can I run Stable Diffusion on 8GB VRAM?
Yes. SD 1.5 and the SDXL base model run natively on 8GB. SD 3.5 Medium, Flux.2 Klein 4B, and Z-Image Turbo fit with FP8 checkpoints. Stacking the SDXL Refiner, multiple LoRAs, or ControlNet usually requires the --lowvram flag.
How much VRAM does SDXL need?
SDXL needs 8GB of VRAM at minimum, according to Stability AI's launch specs. 12GB is comfortable for everyday work, and 16GB covers the Refiner, ControlNet, and stacked LoRAs.
Can I run Stable Diffusion without a GPU?
Yes, with backends like OpenVINO, but CPU generation is dramatically slower than GPU generation. A single image can take tens of minutes depending on the CPU and model, making iterative work impractical. A cloud GPU is the more viable alternative.
How do I fix CUDA out-of-memory errors in Stable Diffusion?
Reduce the output resolution first, then add the --lowvram flag if errors persist. Switching from FP16 to FP8 cuts VRAM usage by 40-50%. GGUF-quantized models bring larger models within reach of 6-8GB cards.