Icon

Updated weekly

OW

OW

Icon

Updated weekly

OW

Output Weekly

Output Weekly

Tools, workflows and honest rankings for people who make things with AI

Tools, workflows and honest rankings for people who make things with AI

Platforms

7 best places to run Stable Diffusion in 2026

Danielle Kerr

A computer with a blue light

Key takeaways

  • ComfyUI is the first pick. It's free under the GPL-3.0 license, lists SD 1.5, SDXL, and SD 3.5 among its native models, and has a desktop app for Windows and macOS.

  • Mage is the hosted pick. Its docs say SD 1.5, SDXL, and SD 3.5 Large run on its GPUs with nothing to install, and all three are unlimited from the $10 Basic plan at base settings.

  • Stability AI's newest Stable Diffusion image model is still SD 3.5, released in October 2024. Its Community License is free for commercial use up to $1M in annual revenue.

  • InvokeAI lists 8GB of VRAM as the minimum for SDXL and 4GB for SD 1.5.

  • Renting a GPU on RunPod starts at $0.89 an hour for an RTX 4090 on Secure Cloud, and the Stability AI API charges per generation.

What's the best place to run Stable Diffusion?

For most people with a recent graphics card, ComfyUI on your own machine is the first place to try, since it's free and supports SD 1.5, SDXL, and SD 3.5. If you don't have the hardware or don't want to install anything, a hosted option like Mage or the Stability AI API gets you generating in minutes.

"Stable Diffusion" covers a family of models, so the right place depends on which one you want. Our explainer on Stable Diffusion covers the lineage from SD 1.5 to SDXL to SD 3.5, plus the community checkpoints built on them. We ranked each place on which of those models it runs, what it costs to start, what hardware or account it needs, how its license terms read, and how much control it gives you.

The seven places at a glance

Place

Type

Stable Diffusion models

Starting cost

Needs

ComfyUI

Local app

SD 1.5, SDXL, SD 3.5

Free

Windows, Linux, or macOS

Mage

Hosted web

SD 1.5, SDXL, SD 3.5 Large, Pony, Illustrious

Free account, $10/mo for unlimited

Browser

InvokeAI

Local app

SD 1.5, SDXL, SD 3.5

Free

8GB VRAM for SDXL

Forge and AUTOMATIC1111

Local web UI

SDXL, SD3 (AUTOMATIC1111)

Free

Python 3.10.6 on Windows

Stability AI API

Hosted API

SD 3.5 Large, Large Turbo, Medium

$0.035 to $0.065 per image

API key and code

RunPod

Cloud GPU

Whatever you install

$0.89/hr for an RTX 4090

Account and credits

Draw Things

Local app

SDXL

Free with in-app purchases

iPhone, iPad, or Mac

1. ComfyUI: best for control and model coverage

ComfyUI is an open-source, node-based interface that runs on your own machine. You build each generation as a graph of connected boxes.

Strengths: The README lists SD 1.5, SDXL, and SD 3.5 under "Broad native model support," alongside FLUX, Qwen Image, and video models. It runs on Windows, Linux, and macOS, and Comfy Desktop is "the recommended way to install ComfyUI" on Windows and macOS. The most recent release, v0.46.0, shipped on November 18. The README also says it can run large models on as little as 4GB of VRAM with 8GB of RAM. The project sells a hosted version too, Comfy Cloud, with a free tier and a $20-a-month Standard plan on monthly billing.

Limits: The graph takes longer to learn than a single prompt form, and you supply the hardware or pay for Comfy Cloud. Our Windows setup guide walks through the install.

Best for: People with an NVIDIA, AMD, or Apple Silicon machine who want the widest model support and control over every step.

2. Mage: best for a hosted setup with community checkpoints

Mage is a browser-based image and video platform. Its docs say Stable Diffusion 1.5, SDXL, and SD 3.5 Large "run on Mage's GPUs with nothing to install," and the Illustrious and Pony families run the same way as SDXL checkpoints.

Strengths: At base settings, all three base models are unlimited from Basic ($10 a month). One LoRA runs on Basic, two on Pro ($30), and more on Pro Plus ($60). ControlNet on SD 1.5 and SDXL starts at Pro. Pro Plus and Max can import checkpoints and LoRAs from CivitAI, Hugging Face, or Google Drive links. Mage describes its policy as "creative freedom within clear, lawful boundaries."

Limits: SD 3.5 Large takes no LoRAs, and the docs say there's "no shell access, custom node graph, or arbitrary extension."

Best for: Anyone who wants community LoRAs and checkpoints in a browser at a flat monthly price.

3. InvokeAI: best for an app-style local install

InvokeAI is an open-source creative engine with a web-based interface, installed through a launcher.

Strengths: Supported models run from SD 1.5, SD 2.0, SDXL, and SD 3.5 Medium and Large to FLUX, Z-Image, and Qwen Image, per the repository. It runs on Windows 10 and later, macOS 14 and later, and Linux. GitHub lists Apache-2.0 and three other licenses, and the README calls it "free to use under a commercially-friendly license."

Limits: Its requirements table gives SD 1.5 a minimum of 4GB of VRAM and 8GB of RAM, and SDXL 8GB of VRAM and 16GB of RAM on an Nvidia 20-series card or newer. Apple Silicon Macs work, but it recommends 16GB or more of memory. Our VRAM guide covers what other models ask for.

Best for: People who want a launcher install and a conventional interface instead of a node graph.

4. Forge and AUTOMATIC1111: best for the classic web UI

AUTOMATIC1111 is a Gradio-based web interface under the AGPL-3.0 license, and Forge is a platform built on top of it.

Strengths: AUTOMATIC1111's README lists txt2img and img2img modes, inpainting, outpainting, LoRA support, textual inversion, and an API. SDXL support arrived in version 1.5.0, and release 1.10.0 added SD3. Forge is based on version 1.10.1 and adds GPU weight and offload controls, plus integrated ControlNet. The AUTOMATIC1111 wiki describes low-memory modes meant to make 512x512 images possible on 4GB cards.

Limits: The Forge README says it's "under some constructions," and it syncs with AUTOMATIC1111 about every 90 days. AUTOMATIC1111 recommends Python 3.10.6 on Windows and warns that newer Python versions don't support torch.

Best for: Users who prefer a form-style interface with modes for text-to-image and image-to-image.

5. Stability AI API: best for building SD 3.5 into an app

Stability AI's developer platform serves Stable Diffusion 3.5 over an API, so no graphics card is involved.

Strengths: SD 3.5 Large costs 6.5 credits, Large Turbo 4, and Medium 3.5, and 1 credit equals $0.01 on Stability's pricing page. That works out to $0.065, $0.04, and $0.035. The weights are under the Stability AI Community License, which covers any business making under $1M a year. Above that, Stability points to an Enterprise License.

Limits: You need an API key and some code, or a client app that takes one. The pricing page says "Pricing is subject to change as we improve our models and infrastructure."

Best for: Developers who want SD 3.5 on demand and pay per image.

6. RunPod: best for renting a GPU by the hour

RunPod rents GPU servers, and its docs include a ComfyUI tutorial.

Strengths: On Secure Cloud, an RTX 4090 rents for $0.89 an hour, an RTX 5090 for $1.39, and an A40 for $0.59, billed per second, according to RunPod's pricing. The docs offer a Pod with ComfyUI pre-installed, plus a Serverless route for building an endpoint. A network volume keeps models between sessions.

Limits: The tutorial asks for at least $10 in credits, warns a Pod can take up to 30 minutes to initialize, and says the template includes no models. It also says to stop your Pod when you finish "to avoid unnecessary charges."

Best for: Anyone without a strong local GPU who wants ComfyUI on rented hardware for big models or batch runs.

7. Draw Things: best for iPhone, iPad, and Mac

Draw Things is an app from Draw Things, Inc. that generates images on the device.

Strengths: The developer's App Store listing promises "Create art 100% offline" and names SDXL, FLUX.1, Qwen Image, Z Image, and FLUX.2 among the supported models. It also lists on-device LoRA training, ControlNet, and inpainting, and lets you import community models and LoRAs. The app's code is on GitHub under GPL-3.0.

Limits: It's free with in-app purchases, including a $12.99 Draw Things+ option. The listing names SDXL but not SD 1.5 or SD 3.5. It needs iOS 15.4 or later on iPhone and iPad, or macOS 12.4 or later on a Mac.

Best for: Apple device owners who want offline generation. Our Mac vs. PC guide covers the trade-offs.

How to choose a place to run Stable Diffusion

  • If you own an NVIDIA card with 8GB of VRAM or more, start with ComfyUI or InvokeAI. Our GPU guide covers cards if you're buying.

  • If you're on a Mac, try Draw Things or ComfyUI Desktop.

  • If you have no graphics card, Mage gives a flat monthly price, the Stability AI API charges per image, and RunPod bills by the hour.

  • For community checkpoints like Pony and Illustrious, read how they differ from SDXL, since each has its own license terms.

Running Stable Diffusion FAQ

Is Stable Diffusion free to run?

The local tools here are free, and so are the model downloads, though SD 3.5 requires accepting a license. Its Community License is free for commercial use under $1M in annual revenue. Hosted options like Mage, the Stability AI API, and RunPod charge for the compute.

How much VRAM do I need?

InvokeAI's table lists 4GB for SD 1.5 and 8GB for SDXL. Stability AI says SD 3.5 Medium needs 9.9GB of VRAM, not counting the text encoders. Larger models and offloading change those numbers, so check our VRAM guide.

Can I run Stable Diffusion on a Mac?

Yes. ComfyUI has a macOS desktop app, InvokeAI says all Apple Silicon Macs work, and Draw Things runs on Mac, iPhone, and iPad. Speed and add-on support differ from NVIDIA PCs, which our Mac vs. PC guide covers.

Which Stable Diffusion version should I start with?

SD 3.5 is Stability AI's newest image model, and its news page lists no newer one in 2026. SDXL has a much larger library of community checkpoints, and most of the places in this list run it. The Stability AI API is the exception, since it serves the SD 3.5 family.

Can I use my own LoRAs and checkpoints?

On local tools, yes, since you load the files yourself. Mage lets you stack LoRAs on SD 1.5 and SDXL, and Pro Plus and Max can import checkpoints. The Stability AI API runs Stability's own models.