I tested the most popular AI image generators to discover their greatest strengths and weaknesses.
At Ahrefs, we have a team of extremely skilled (and very human) designers, but not everyone has that luxury. I wanted to know: are AI image generators useful for spinning up quick social media posts, creating blog post graphics, or saving a few bucks on expensive stock photography?
So I tested out the most popular cloud-based text-to-image tools: DALL-E 3 (available in ChatGPT), Midjourney, Canva’s Magic Media, Adobe Firefly, and the very new Gemini for Workspace.
All these tools generate images in a few clicks, without needing to do anything complicated like training custom models or running programs locally on your computer.
The best AI image generator is, in my opinion, Adobe Firefly. All the models had their own strengths, but Firefly offered maximum control over image generation and image editing.
Here are the pros and cons (and many, many images) sharing my experience with each.
Adobe Firefly
Best for maximum control over images
25 free credits per month; $4.99 for 100 credits
Midjourney
Best for beautiful images
From $10/m for 200 generations
DALL-E 3 / ChatGPT
Best for data visualization
2 free images per day on the Free plan; full access starts at $20/m on the Plus plan
Canva Magic Media
Best for generating vector images
50 images available for Canva Free users; 500 images per month for paid users (from $14.99/m)
Gemini for Workspace
Best for quick concepting
Available as a Google Workspace add-on from $20/m
ChatGPT, DALL-E 3—the image generation model offered as part of ChatGPT—will be most people’s first introduction to AI image generators. That’s a shame, because it’s one of the worst.
Look for a moment and you’ll spot nonsense text, furniture blending into the background, a weird uncanny-valley glow to the main character, straight lines that are never straight… and most of ChatGPT’s images suffer from the same issues.
Here’s ChatGPT trying to gaslight me into believing that this is a photograph of a home office (the trees look like a freaking pointillism painting):
These issues are at least less obvious in cartoon imagery. Here’s our character holding a key again:
Not bad, despite a couple of AI-isms, like the double-ended key and weird abstract backpack charm. Unfortunately, I couldn’t remove these little quirks, because even though ChatGPT recently added the ability to highlight parts of the image to selectively edit, this feature was super unreliable when I tested it.
On one occasion, ChatGPT even decided that, actually, no, it didn’t want me to do any image editing:
Without much control over image generation or editing, DALL-E 3 is a bit of a crapshoot, and it’s virtually impossible to carry consistent styles across images.
When I tried to make a new image with the same cartoon character, it changed style radically:
You can’t easily upscale your images either, and when I asked ChatGPT to resize a YouTube thumbnail to 16:9 resolution, it decided to write a Python script to stretch the image to landscape format.
Which, err… did not look good:
When I tried to refine the prompt to reflect Ahrefs’ brand guidelines, it gave me a lecture on designing thumbnails, and didn’t actually make an image.
Generating images with ChatGPT reminds me playing the video game DOOM on a calculator. It might technically be possible, but you probably shouldn’t do it.
ChatGPT had one big redeeming virtue, where its penchant for Python was extremely useful: data visualization. It was the only AI image generator capable of actually turning a list of data points into an accurate graph:
And it can handle more complex data visualisations too:
This is a different type of “image generation”, but for someone like me who wrangles data on a daily basis, incredibly useful, and a feature I use all the time.