Stable Diffusion WebUI

Last reviewed Aug 26, 2026 · 789 words

Buying advice first: AUTOMATIC1111's Stable Diffusion WebUI is no longer the default I'd hand a newcomer, because releases have slowed to a crawl while the image-generation world kept moving. It is still the right choice in one specific situation — you want a conventional point-and-click interface with the largest back catalogue of extensions, tutorials, and forum answers ever assembled around local image generation, and you're happy running mostly SD 1.5 and SDXL-era models. That situation describes a lot of people. It just no longer describes everyone, and pretending 164,000 GitHub stars means active momentum would be doing you a disservice.

VRAM decides what you can run; the 8 GB RAM figure doesn't

System RAM barely matters here — the catalogue's 8 GB minimum is about the Python process and model loading, and any modern machine clears it. The number that actually gates your experience is GPU VRAM. Rough working figures: SD 1.5 models run comfortably on 4 GB of VRAM with the --medvram flag, and well on 6-8 GB without it. SDXL wants 8 GB as a practical floor and is happier with 12. CPU-only generation technically works and is measured in minutes per image; treat it as a demo, not a workflow. NVIDIA remains the smooth path — AMD via ROCm and Apple Silicon via PyTorch's MPS backend both function, with more setup friction and fewer working extensions. Before buying hardware for this, run the numbers in VRAM math.

The install is one script and one very specific Python

Setup is webui.sh on Linux/macOS or webui-user.bat on Windows: the script creates a venv, pulls PyTorch and roughly 2 GB of dependencies, and serves a Gradio UI on port 7860. The classic trap is the Python version. The project targets Python 3.10, and pointing it at whatever newer Python your distro ships is the single most common cause of a broken first install. Install 3.10 alongside your system Python and set python_cmd in webui-user.sh before you start. Docker images from the community sidestep the version problem at the cost of GPU passthrough config. Two flags worth knowing from day one: --xformers for a real memory-and-speed win on NVIDIA, and --listen if you want the UI reachable from other machines — and if you set the latter, put authentication in front, because the UI ships with none.

Checkpoints, LoRAs, and where files actually go

The UI is only as good as the models you feed it. Checkpoints go in models/Stable-diffusion/, LoRAs in models/Lora/, VAEs in models/VAE/, and the UI picks them up on refresh without a restart. Expect checkpoints at 2-7 GB each; a working library eats 50 GB of disk before you notice. The extension ecosystem is the feature no rival has fully replicated: ControlNet for pose and composition control, regional prompting, upscaler pipelines, training helpers — installable from a URL inside the UI. Extensions are also arbitrary code running with the app's permissions, so treat installing one like running any stranger's script, and expect that some older ones no longer work against current code.

The maintenance question, and the exits

The repo has gone long stretches without a release, and support for the newest model generations tends to arrive late or via extensions rather than in core. Two exits are worth knowing before you commit. ComfyUI is the actively developed power option — a node-graph interface that is genuinely harder to learn and genuinely more capable, especially for new model architectures and complex pipelines; the full trade-off is in AUTOMATIC1111 vs ComfyUI. The Forge fork keeps the familiar A1111 interface on a reworked, faster backend, and most A1111 knowledge transfers to it directly. Your models are just files on disk, so switching UIs later costs you an afternoon, not your library.

What I'd do

If you have an NVIDIA card with 8 GB or more and want the gentlest on-ramp into local image generation, install A1111 (or Forge for the same interface with better performance), grab one well-regarded SD 1.5 checkpoint and one SDXL checkpoint, and lean on the four years of accumulated guides — that stockpile of answers is the real product. Add ControlNet in week one; it is the extension that turns prompting into directing. But go in with eyes open about the release cadence, keep your models organized so they're portable, and when you hit the ceiling — a new model family core doesn't support, a pipeline extensions can't express — move to ComfyUI rather than waiting for the old champion to catch up.