What’s the most popular diffusion model?

But first, what are they?

Before we start comparing diffusion models and finding out which is the most popular, let’s go over what they are.

Diffusion models are generative AI models that create content like images, video and audio. They typically do this in two separate steps (as in the first step takes place independently of the second):

1. Adding noise: imagine you’ve got a clear photo, or “training data”. This could be an image of a cat, a tree, whatever. Now, imagine scribbling all over it. You keep adding more and more scribbles until you can’t even see the original picture anymore. The purpose of this is to give the AI a ‘map’ that allows it to ‘get back’ to the original image.
2. Removing noise: once the original image is completely covered by noise, the model uses the ’map’ to remove the noise, step by step. The trick is to train the model to get back to the original image before all that noise was added. It practices on many examples, and over time, it gets better at figuring this out.

Once trained, the model starts with a random mess (like a blank canvas of noise) and uses what it’s learned about cleaning up noise to create a new image from scratch.

Think of it like a landscape painter at work. They don’t copy any specific image they’ve seen. Instead, they use what they’ve learned to paint a new scene from scratch. Each stroke and colour choice is influenced by their internalised knowledge. The result? A unique, but realistic painting.

Let’s cut it down to five

There are loads of popular diffusion models out there. But to keep things simple, we’ve picked our top five.

1. Midjourney

Midjourney is an independent research lab’s brainchild.

Here’s what one of our designers, Nikolas Falas, thinks about it:

“From a graphic designer’s perspective, I personally believe that Midjourney is the best AI image generation tool. I have tried both Adobe Firefly and DALL-E, but nothing compares to Midjourney’s attention to detail. It allows you to customise the design output extensively when using the right prompts, in a level that you can even use it as a tool for storyboarding. 

“From what you can see in the images below, Midjourney is great at combining real-world images with fantastical concepts. That’s why it has been used by a lot of artists to create some amazing surrealistic art. Sara is a great example of this.

“As it’s a standalone software, it took some time to familiarise myself with it and especially how I come up with the right prompts to get the results that I was aiming for (a copywriter friend of mine helped me with the prompts as he has been using it too). Finally, even with the similar prompts, it has been difficult to create visuals that have the same look and feel. It’s an amazing tool, but it requires a lot of time to get it right.”

2. DALL-E

OpenAI’s DALL-E 3 is built natively on, and integrated with, ChatGPT. It’s designed to create images perfectly in line with your text prompt and the integration means images can be generated in a conversational way. This makes it more accessible. The idea is that a lot can be achieved with natural language descriptions.

3. Imagen

Google’s thrown its hat in the ring with Imagen. Imagen 2 is currently available. Imagen 3 is on the way, but currently only available to select creators.

Tom Pallot, our Head of Marketing, says:

“These models all bring different things to the table so it’s valuable (and fun) to be able to experiment with them.

“One of Imagen 2’s strengths is being able to generate photo realistic, lifelike images. Google says this is down to the natural distribution of its training data vs using pre-defined styles. You can see the difference in these two examples here that use the same prompt: ‘a photograph of a canoe on a gentle stream’. The first one, generated by Imagen 2, is clearly more convincing as a photo.

“We can expect new, more powerful, exciting models to keep being released. Beyond image gen too – lots of new models, that do lots of useful things for your business. The most important thing is giving your team the best tools for the job in the most cost-effective way.”

Made using Imagen 2

Made using DALL-E 3

4. Firefly

Adobe Firefly has come in handy for our designers, like Red Howell:

“Firefly has become a regular part of our workflow, particularly within Adobe Photoshop. The generative tools have been easy for the creative team to master, sometimes requiring little to no prompting thanks to Adobe’s large library of licensed content powering Firefly’s training.

“We’re now able to stick much closer to our original creative visions for clients and we can compromise less when trying to achieve certain aesthetics, particularly when it comes to the use of imagery. For example, great photos that previously may have been unusable for certain aspect ratios or print formats can now be easily adjusted or expanded on with less time on the tools and more focus on the overall vision.”

5. Stable Diffusion

Stability AI’s latest offering is Stable Diffusion 3. They’re saying it’s top-notch for photorealism, handling tricky prompts and creating clear text.

So, we’ve picked these five models to focus on. Next, we looked at how often people search for them in some of the world’s tech hotspots:

• Japan
• United States
• South Korea
• Germany
• Singapore
• United Kingdom
• Sweden
• Israel
• Switzerland

We used Google’s Keyword Planner to check the average search volume over the last year for each tool, in each country.

*See notes on our workings at the bottom.

Midjourney’s not just winning – it’s dominating

Midjourney comes out on top in 8 out of 9 countries we checked. Search volumes for ‘Midjourney’ were miles ahead compared to the other diffusion models – 905,600 average monthly searches to be exact across the countries we checked.

Sad times for Google – Imagen came in dead last, scraping together 121,810 average monthly searches. Tbh, ‘Imagen Google’ did better, but we stuck with ‘Imagen’ to keep things fair.

We’ve broken down and ranked the popularity of each diffusion model per country.

Japan 🇯🇵

United States 🇺🇸

South Korea 🇰🇷

Germany 🇩🇪

Singapore 🇸🇬

United Kingdom 🇬🇧

Sweden 🇸🇪

Israel 🇮🇱

Switzerland 🇨🇭

What’s next for diffusion models?

Our AI Director, Luke Budka, says:

“The diffusion model market is heating up. Canva just bought Leonardo, and there are new services being launched all the time – e.g. iStock, a Getty company, just launched a new AI Generator. It kindly doesn’t let you sell the images you create, but it does use anything you create for future training, thanks guys. AI companies are also fighting for training data resource: Figma’s the latest to announce that it’s training its models on its users’ creations.

“As always, make sure you’re reading the T&Cs and use the best models for the job. Expect text to video to take the hot seat between now and the end of the year as the market eagerly awaits OpenAI’s Sora model.”

*Some of our workings

We’re not saying this is perfect science, but it gives us a good sense of the market, nevertheless. Here’s how we went about it.

We searched “[name of AI tool]” to gauge the amount of people searching for each one to determine which was the most popular.

BUT some things to consider (which affect the search volume):

• Firefly is also a virtual learning platform for schools around the world and a Malaysian airline.
• Imagen is also an AI photo editing product, and a cloud-native media asset management platform from Reuters.