Stable Diffusion
Overview
Stable Diffusion is an open text-to-image model that runs on your own GPU. Nothing leaves your machine — no prompts, no images — which means no usage fees, no content filters and no rate limits. The current open family published by Stability AI is Stable Diffusion 3.5 (Large / Large Turbo / Medium), while the deepest pool of community checkpoints and LoRAs is still built on SDXL. This page focuses on how to actually use Stable Diffusion as a tool (it is not an overview of the Stability AI company). In practice the frontend is what matters day-to-day, and the three dominant options are AUTOMATIC1111 (A1111), ComfyUI and Forge.
Key Features
Local execution, full control
Models, LoRAs, ControlNets, VAEs and embeddings all sit on your disk. Prompts and outputs never reach the internet, and NSFW / trademark handling is entirely up to you.
Frontend choices
- AUTOMATIC1111 (A1111): classic form-based WebUI, beginner-friendly, serves at http://127.0.0.1:7860.
- ComfyUI: node-based visual workflow canvas. Pro users are migrating here. Serves at http://127.0.0.1:8188.
- Forge: A1111-compatible fork with automatic memory management, making SDXL feasible even on 6GB VRAM cards.
Huge extension ecosystem
Communities on Civitai and Hugging Face ship checkpoints, LoRAs, ControlNets and IP-Adapters you can mix and match for very fine control over characters, styles and composition.
Licensing and commercial use
SD 3.5 weights ship under the Stability Community License, which is free for individuals and companies below a revenue threshold; above it you need an enterprise license. If you would rather not run your own GPU, Stability's hosted API (Stable Image Ultra and friends) is the alternative.
Plans & Pricing
| Setup | Cost | Notes |
|---|---|---|
| Local | Free (power only) | GPU required |
| Cloud GPU (e.g. RunPod) | Hourly | Use when VRAM is low |
| Stability API | Pay-as-you-go | For running without your own GPU |
| Models themselves | Free download | Respect each model's license |
8GB VRAM is recommended for SDXL. With --lowvram or Forge's memory manager, even 6GB cards can run it. SD 3.5 Large asks for considerably more.
Typical Use Cases
- Training your own LoRA for a specific character or product
- Pose / composition lock-in with ControlNet
- Running private in-house models on internal data
- Batch generation and frames for animation/video pipelines
Strengths
- Fully local — prompts and images stay on your machine
- Zero usage fees, unlimited generations
- Surgical control via LoRA / ControlNet
- Massive community assets (Civitai etc.)
Weaknesses
- Requires a capable GPU
- Non-trivial initial setup
- Out-of-the-box quality now trails Midjourney / Flux / Ideogram in most comparisons
Official Info
- Models: Hugging Face
- A1111: https://github.com/AUTOMATIC1111/stable-diffusion-webui
- ComfyUI: https://github.com/comfyanonymous/ComfyUI