Hosted Stable Diffusion services compared — Speeding up AI image generation

If you even remotely follow the news in the creative AI space, you’ll have come across services like MidJourney, from the 11 full-time staff team led by David Holz and DALL·E 2 by OpenAI.
Now Adobe is also on the playing field with their Firefly software and generative AI built right into Photoshop.

However, we as artists have long yearned for more control over the processes that run these diffusion models, that control is available via open source apps to interface with the diffusion models, most popular which include InvokeAI and Automatic1111. Both are built on Gradio, an interface builder that quickly allows developers to spin up a user interface for what would normally be command-line tools.

Now if you’re anything like me you are on a Mac or a Windows laptop without a big gaming GPU in your rig. You’ll potentially have tried out some of the native ways to interact with Stable Diffusion and some of the other hosted competing services mentioned before. Through that process you’ll have realized that in order
to gain some kind of control over this technology and in order to be able to use it in production, you need the speed to be able to iterate and learn.

You can’t be waiting more than a minute for each image generation to complete.
This is the gap that independent services like the ones we’re gonna be looking at today, fills. They allow you to interact with these extremely powerful diffusion models
without having a massive GPU locally on your computer. Like in my use-case, on a Mac M1 laptop.

In today’s overview, I will be focusing specifically on automatic 1111s UI, but most of these services also allow you to run Invoke AI, which is the more user-friendly, less technical option.

ComfyUI’s node view

A note on ComfyUI: A while ago I saw the initial release of ComfyUI which is a node-based way of interacting with the diffusion models. This is naturally extremely interesting as this allows for complex workflows to be automated and for ultimate control of the process. However, the open source implementation is not quite there yet for the regular user and I have yet to see a service offering a hosted version so you would need to have a local GPU for this.

Another interesting aspect pf a hosted service is the addition of the functionality that enables you to run the service in API mode in order to use the clouds computing power on your local computer for example via the Photoshop plugin.

In today’s review, I will be looking at three hosted auctions and do a couple of test renders on each to compare price and speed both for batches and for single image generations.This is far from a scientific, rigorous process, but should give you an idea of the kind of performance that you will be able to see on the individual service.

I will also be noting my user experience with each of the services, and how quickly I was able to spin up a version of the UI and generate my first images with a custom model, custom LoRa and lots of custom settings. offers an extensive selection of preloaded models and provides the capability to add a new model directly on the page. What’s nice about the system structure is the ability to add models and LoRAs before server initialization, this means you won’t waste time configuring models and downloading them once the service is live.

During the initial setup, it took me about 7 minutes to download files from the automatic 1111 GitHub. However, I faced some confusion initially as my models and LoRAs weren’t immediately available in the interface. To resolve this, I had to consult the FAQ section and activate each of them. Though this did result in some error logs, a quick server restart allowed me to proceed with the models and LoRAs at my disposal.

For the service, it’s important to remember that model loading is required each time, unless you follow the instructions in the FAQ to modify the existing settings in the config.json file. After gaining a basic understanding of the service, and performing model loading and restarting, I was 23 minutes into the process with the models and LORAs loaded and ready for generation.

In testing, I found that my example images (identical across all three services tested) were generated in an impressive 5.49 seconds. Additionally, generating 10 batches of images consistently took 58 seconds. offers its pricing in two forms. Firstly, a slower pay-as-you-go model where the rate is 50 cents per hour, and secondly, a prepaid plan that offers 30 hours for $19, after which the 50 cents per hour rate applies. The Pro plan offers better hardware, claiming to generate images three times faster, which was the only option I tested.


Run Diffusion, which might be the most well-known in this field, is a prominent and popular way to start with automatic 1111 as a hosted service.

The configuration process isn’t as straightforward. To start, you need to log in to the service and configure your session. But be aware that you can’t load custom models unless you’re a member of the “Creators Club”. For newcomers, differentiating between individual plans can also be challenging. They recently reduced the price of the large model to $1.75 per hour is cool, it is also the only hardware I would recommend, particularly for beginners who want to iterate quickly.

The Creators Club also grants access to more advanced options, like training your own LORAs and Dream Booth models, and providing 100 gigabytes of storage space.

Once you understand the nuances of configuring your session, the process becomes quite simple. Visit and configure your session at the top. For this evaluation, I selected the Automatic UI, the latest large model, and the Creator’s Club to use private models and LoRAs.

The largest hardware plan from RunDiffusion managed to generate a single image in 10.85 seconds, which is almost twice the time taken by Stadio. However, for a batch of 10, it clocked in at a faster 39 seconds. Overall, Run Diffusion is a solid beginner platform for learning and understanding how to create visuals with stable diffusion.

Stable Matic

Stablematic presents one of the easiest methods to start with Stable Diffusion. Setting it up is really straightforward. You simply need to add credits to your account, head over to the ‘launch playground’, and initiate the app.

You can download custom models directly from CIVITAI before starting the server. It’s an efficient process, and the pricing is also reasonable, costing around $1 per GPU hour. I found it quite simple to download a custom model. However, I encountered a bit of difficulty when I tried to add a LoRA on my own. To resolve this, I’ve reached out to the Stablematic team for further clarification.

Moving on to the user experience, I found the interface to be really snappy and responsive. However, generating an initial image took more time compared to other competitors in the market. For instance, it took 49 seconds to create the first image. During a second run, a single image was produced within 20 seconds.

A batch of 10 images was also generated without the LoRA. It took a substantial amount of time, three minutes and one second, to be precise.

Overall, Stablematic is a solid option if you’re not prioritizing the fastest hardware, but rather seeking a straightforward, one-click launch solution for starting to generate images images.

Summary & Comparison table

Service NameSpeed / Single ImageSpeed / Batch of 10Pricing ModelPriceSign-up Link
Stadio.ai5.49 seconds58 secondsPay-as-you-go or Prepaid plan$0.50 per hour or $19 for 30 hours
RunDiffusion10.85 seconds39 secondsPay-as-you-go or Membership (“Creators Club”)Starting at $0.50/hr (without “Creators Club”)
StableMatic20 seconds (average)181 secondsPay-as-you-go$8/month (Public Alpha Pricing)
* Keep in mind that this data is based on a test run on a single day with specific settings. The results might not be indicative of typical performance

The creative AI space continues to expand, offering more control over the processes that run diffusion models with services like MidJourney, DALL·E 2 by OpenAI, Adobe’s Firefly, InvokeAI and Automatic1111. These hosted services fill the gap for artists and users who don’t have a large GPU on their local computer and want to speed up their AI image generation process.

We examined three of these services in this review:, RunDiffusion, and StableMatic. boasts a pre-loaded selection of models (to be fair this is not unique to them) and offers an impressive generation speed of 5.49 seconds per image, and just under a minute for ten images, but the initial setup was somewhat confusing and time-consuming. The pricing is structured as pay-as-you-go or a prepaid plan.

RunDiffusion is a prominent player and provides a somewhat straightforward configuration process once understood. Although its generation speed is slower than Stadio, it offers several benefits for members of their “Creators Club” such as access to advanced options and additional storage space.

Lastly, StableMatic offers a user-friendly setup process and reasonable pricing but has longer image generation times. It’s a solid choice for users who aren’t as concerned about speed but want a simple, one-click launch solution.