How to Run Stable Diffusion Locally (2026)
Run Stable Diffusion and FLUX locally: which interface to use (Forge, ComfyUI, A1111), the real VRAM you need for SDXL, SD3.5, and FLUX, and a clean setup path.
Prerequisites
- ✓A GPU with at least 8GB VRAM (12GB+ recommended)
- ✓About 20GB free disk space for a model and tools
- ✓Windows or Linux
Quick answer
To run image generation locally, install Forge (the most VRAM-efficient beginner UI) or ComfyUI (node-based, for power users), download a model, and generate. VRAM is the gate: SDXL is comfortable on 8GB with Forge, SD 3.5 and FLUX want 12GB at FP8, and FLUX in full BF16 wants 24GB. 12GB is the honest floor and 24GB is the target for the recent models — a used RTX 3060 12GB is the cheapest sensible card.
Running Stable Diffusion (and FLUX, the other major local image model) on your own GPU means unlimited, private, free image generation. The setup is more involved than running a chat model, but it comes down to two choices: which interface you use, and whether your VRAM fits the model you want. This guide handles both.
If you have not picked a GPU yet, the VRAM requirements guide and best GPU for local LLMs guide cover image-generation memory needs too.
How Much VRAM You Actually Need
VRAM is the real constraint for local image generation. Verified targets for 2026:
| Model | FP8 / optimized | Full precision | Notes |
|---|---|---|---|
| Stable Diffusion 1.5 | 4–5GB | 6GB | Runs on almost anything |
| SDXL | 8GB (with Forge) | 12GB | The mainstream baseline |
| Stable Diffusion 3.5 | ~12GB | tight on 12GB | FP8 fits comfortably; FP16 is tight |
| FLUX.1 Dev | ~12GB (FP8) | 24GB (BF16) | The quality leader; hungry |
The honest summary: 12GB is the floor for the models released in the last 18 months, and 24GB is the target if you want to run FLUX or SD 3.5 at full quality without juggling. A used RTX 3060 12GB is the cheapest card that handles SDXL well; for FLUX at full quality, you want 24GB (a used RTX 3090).
Pick Your Interface
There are three mainstream front ends, all free:
- Forge — a fork of AUTOMATIC1111 that uses roughly 30–50% less VRAM for the same quality. The best starting point for most people: a friendly web UI without the memory penalty.
- ComfyUI — a node-based workflow graph. Steeper learning curve, maximum control and flexibility, and it is also very VRAM-efficient (~30–40% less than A1111). What most advanced users run.
- AUTOMATIC1111 (A1111) — the classic web UI. Familiar and well-documented, but less VRAM-efficient than Forge or ComfyUI.
Recommendation: start with Forge if you want the easiest path and the lowest VRAM use, and move to ComfyUI later if you want node-based control over complex workflows.
Step-By-Step Setup
1. Confirm your hardware
Check your VRAM against the table above. An NVIDIA GPU is the lowest-friction choice — the image-generation ecosystem is CUDA-first. AMD works but with more setup (see the NVIDIA vs AMD guide).
2. Install your interface
Download Forge (or ComfyUI) from its official source and follow its install steps. On Windows this is typically a downloadable package; on Linux you clone the repo and run its setup. Both bundle the Python environment they need.
3. Download a model
Get a model checkpoint that fits your VRAM:
- 8–12GB: start with an SDXL checkpoint — the best quality-to-VRAM balance.
- 12GB+: try SD 3.5 (FP8) or FLUX.1 Dev (FP8) for newer-architecture results.
- 24GB: run FLUX.1 Dev at full quality.
Place the model file in your interface’s models folder (Forge and ComfyUI each document the exact path).
4. Generate
Launch the interface, select your model, type a prompt, and generate. Start with default settings and a modest resolution, then tune. If you hit out-of-memory errors, drop the resolution, use an FP8 model, or switch to Forge/ComfyUI for their memory savings.
Common Mistakes
- Trying FLUX in full precision on 12GB. FLUX.1 Dev in BF16 wants 24GB. Use the FP8 version on 12GB, or step up the card.
- Using A1111 on a tight-VRAM card. Forge and ComfyUI use 30–50% less VRAM for the same output — on 8–12GB that is the difference between working and out-of-memory.
- Buying an 8GB card for serious image work. 8GB runs SD 1.5 and SDXL (optimized), but the recent models want 12GB+. Treat 12GB as the floor.
- Choosing AMD for image generation without expecting friction. The ecosystem is CUDA-first; AMD works but with more setup. NVIDIA is the lower-hassle path here.
What To Do Next
- Confirm your card fits the model you want: VRAM requirements guide.
- Need a better GPU for FLUX at full quality? Best GPU for local AI under $1,000 or a used RTX 3090.
- Want to run chat models too? Best local LLM models 2026.
Frequently Asked Questions
How much VRAM do I need to run Stable Diffusion locally?
SD 1.5 needs 4–5GB and SDXL is comfortable on 8GB with Forge. The newer models want more: SD 3.5 and FLUX.1 Dev fit ~12GB at FP8, and FLUX in full BF16 wants 24GB. Treat 12GB as the floor and 24GB as the target.
Forge, ComfyUI, or AUTOMATIC1111 — which should I use?
Start with Forge for the easiest setup and the lowest VRAM use (30–50% less than A1111). Move to ComfyUI if you want node-based control over complex workflows. A1111 is the classic option but less memory-efficient.
Can I run FLUX locally?
Yes. FLUX.1 Dev runs in FP8 on about 12GB of VRAM, or in full BF16 quality on 24GB. It is the current quality leader for local image generation, and the most VRAM-hungry.
What’s the cheapest GPU for local Stable Diffusion?
A used RTX 3060 12GB is the cheapest card that handles SDXL well and meets the 12GB floor. For FLUX at full quality, step up to a 24GB card like a used RTX 3090.
Does Stable Diffusion work on AMD GPUs?
Yes, but with more friction than NVIDIA — the image-generation ecosystem is CUDA-first. AMD works on Linux with ROCm, but for the smoothest setup, NVIDIA is the recommended choice for image generation.
Last updated: June 2026. VRAM figures and interface comparisons reference current community testing and tooling. We do not publish invented benchmarks. Model availability and optimal settings change as new releases land.