I tested the astir fashionable AI representation generators to observe their top strengths and weaknesses.

At Ahrefs, we person a squad of highly skilled (and precise human) designers, but not everyone has that luxury. I wanted to know: are AI representation generators utile for spinning up speedy societal media posts, creating blog station graphics, oregon redeeming a fewer bucks connected costly banal photography?

So I tested retired the astir fashionable cloud-based text-to-image tools: DALL-E 3 (available successful ChatGPT), Midjourney, Canva’s Magic Media, Adobe Firefly, and the precise caller Gemini for Workspace.

All these tools make images successful a fewer clicks, without needing to bash thing analyzable similar grooming customized models oregon moving programs locally connected your computer.

The champion AI representation generator is, successful my opinion, Adobe Firefly. All the models had their ain strengths, but Firefly offered maximum power implicit representation procreation and representation editing.

Here are the pros and cons (and many, galore images) sharing my acquisition with each.

I wanted to trial each AI representation generator successful a scope of antithetic scenarios, truthful I created tons of prompts crossed 3 main categories:

• Stock photography (e.g. “Stock photograph of a beauteous minimalist location bureau with a presumption of trees outside”)
• Graphics and illustrations (e.g. “A cartoon quality with ginger hairsbreadth carrying a immense aureate cardinal to correspond ‘keyword research’”)
• Data visualizations (e.g “Graph of website postulation data: January 946, February 1071, March…”)

I tested antithetic levels of punctual complexity, but kept my prompts mostly simple. The full constituent of these text-to-image tools is to picture thing that you privation and person the AI make it for you, truthful I purposefully avoided PhD-level punctual engineering oregon nonrecreational plan lingo.

Here’s a photograph of maine moving these tests:

I past judged each AI representation generator’s output crossed a fewer cardinal dimensions:

• Accuracy: how good did the representation generator travel my direction?
• Ease of editing: how casual was it to edit and refine the output?
• Uncanniness: did the output look weird oregon evidently AI-generated?
• Legibility of text: how good did the exemplary grip substance generation?
• Consistency: could I reproduce akin images connected aggregate occasions?
• Usefulness: could I really usage the output successful real life?

Here are my findings.

Adobe Firefly has—by far—the champion editing controls of the representation generators I tested. This isn’t surprising, considering that Adobe makes Photoshop, and Illustrator, and Lightroom, and dozens of different market-leading plan tools.

Here’s an example. The punctual “A cartoon quality with ginger hairsbreadth carrying a immense aureate cardinal to correspond ‘keyword research’” generated a bid of okay-but-not-great images. But successful a fewer clicks, I was capable to hole the biggest problems and dramatically amended the result.

Here’s the before:

In a fewer minutes utilizing Firefly, I was able to:

• Resize the facet ratio from 1:1 to 4:3 utilizing generative fill.
• Fix a missing hand by prompting Firefly to regenerate that circumstantial information of the image.
• Upscale the small, low-quality image to a overmuch much utile 2k resolution.

And here’s the after:

Adobe Firefly besides gives you a ton of power implicit the image-generation process. A large plus: you tin usage existing images arsenic benignant and creation references, making it overmuch easier to make a bid of images with a cohesive style.

Here’s the punctual “A cartoon quality with ginger hairsbreadth carrying a immense magnifying solid to correspond ‘competitor research’”, but utilizing my erstwhile representation procreation arsenic a reference:

The benignant is somewhat different, but they consciousness recognisably similar. You tin besides specify peculiar notation styles, compositions, contented types (like creation versus photo), and adjacent effects (colour, lighting, bokeh, camera angles, you name it).

That means you tin usage the aforesaid punctual but get precise antithetic results. Here’s the effect for the punctual “Beautiful minimalist location bureau with a presumption of trees outside” erstwhile I’ve specified golden hr lighting and warm tones:

And present I’ve utilized the aforesaid punctual but asked for low lighting and cool tones for a precise antithetic vibe:

And due to the fact that Firefly is made by Adobe, you tin import your generated images into different Adobe products to adhd substance oregon edit further. Pretty handy.

Midjourney is beautiful. I’ve been a paying Midjourney lawsuit for 3 years for the elemental crushed that everything it generates is gorgeous, and much aesthetically pleasing than immoderate different AI exemplary I’ve tested.

I usage Midjourney to exemplify my originative writing, and it excels astatine fantasy-style illustration. Here’s an representation I created for 1 of my novels, with nary editing oregon manipulation:

It’s besides beauteous useful for photorealism too. Here’s the punctual “Stock photograph of a beauteous minimalist location bureau with a presumption of trees outside”:

There are a mates of AI-isms (how galore wheels does that seat have?!), but I privation to forgive them due to the fact that the photograph is truthful damn beautiful.

Here’s “Stock photograph of a thoughtful idiosyncratic successful a gathering astatine a bundle company”, featuring an AI-generated antheral truthful handsome I didn’t privation to look successful a reflector for the remainder of the day:

Even Midjourney cartoon illustrations look classy, and astir bully capable to beryllium plucked from the frames of a Pixar film:

Midjourney does person weaknesses. It categorically cannot bash information visualization. Feed it adjacent elemental information and it volition make nonsense (but it volition astatine slightest beryllium beautiful nonsense):

Midjourney’s editing workflows are overmuch amended than they utilized to be, but inactive not precise sophisticated. As good arsenic generating 4 images for each prompt, you person the enactment to:

• Vary immoderate azygous image, either strong or subtle (basically regenerate an representation that’s precise akin to the previous).
• Upscale images you like to higher resolution.
• Remove parts of the image (but not specify what you’d similar to regenerate it with).
• Change the facet ratio (square, 4:3, 16:9, and so on).

Here’s an illustration of varying an image. There are small, subtle differences betwixt each photo, similar the fig of wheels connected the chair—helpful for minimizing immoderate weird AI-isms successful images you like:

These options are obscurity adjacent arsenic precise arsenic Adobe Firefly’s editing workflow, but fixed Midjourney’s quality to marque mostly beauteous images from simple, azygous prompts, this workflow creates amazingly utile images.

(And arsenic a last bonus, you nary longer person to trust connected a janky Discord server to make images—Midjourney’s web app works very well.)

Given the popularity of ChatGPT, DALL-E 3—the representation procreation exemplary offered arsenic portion of ChatGPT—will beryllium astir people’s archetypal instauration to AI representation generators. That’s a shame, due to the fact that it’s 1 of the worst.

To marque this point, here’s what happened erstwhile I asked for a “Stock photograph of idiosyncratic moving connected their laptop successful a New York java shop”:

This is beauteous typical of DALL-E 3: astir of its images look and consciousness similar they’re AI-generated.

Look for a infinitesimal and you’ll spot nonsense text, furnishings blending into the background, a weird uncanny-valley glow to the main character, consecutive lines that are ne'er straight… and astir of ChatGPT’s images endure from the aforesaid issues.

Here’s ChatGPT trying to gaslight maine into believing that this is simply a photograph of a location bureau (the trees look similar a freaking pointillism painting):

These issues are astatine slightest little evident successful cartoon imagery. Here’s our quality holding a key again:

Not bad, contempt a mates of AI-isms, similar the double-ended cardinal and weird abstract backpack charm. Unfortunately, I couldn’t region these small quirks, due to the fact that adjacent though ChatGPT precocious added the quality to item parts of the representation to selectively edit, this diagnostic was ace unreliable erstwhile I tested it.

On 1 occasion, ChatGPT adjacent decided that, actually, no, it didn’t privation maine to do any representation editing:

Without overmuch power implicit representation procreation oregon editing, DALL-E 3 is simply a spot of a crapshoot, and it’s virtually intolerable to transportation accordant styles crossed images.

When I tried to marque a caller representation with the aforesaid cartoon character, it changed benignant radically:

You can’t easy upscale your images either, and erstwhile I asked ChatGPT to resize a YouTube thumbnail to 16:9 resolution, it decided to constitute a Python script to agelong the representation to scenery format.

Which, err… did not look good:

When I tried to refine the punctual to bespeak Ahrefs’ marque guidelines, it gave maine a lecture on designing thumbnails, and didn’t really marque an image.

Generating images with ChatGPT reminds maine playing the video crippled DOOM connected a calculator. It mightiness technically beryllium possible, but you astir apt shouldn’t do it.

ChatGPT had 1 large redeeming virtue, wherever its penchant for Python was highly useful: information visualization. It was the only AI representation generator susceptible of really turning a database of information points into an close graph:

And it tin grip much analyzable information visualisations too:

This is simply a antithetic benignant of “image generation”, but for idiosyncratic similar maine who wrangles information connected a regular basis, incredibly useful, and a diagnostic I usage each the time.

Canva’s Magic Media is an AI representation generator embedded straight wrong the main Canva app. To get started, you’re offered a prime of image, graphic, oregon video.

It handles banal photography beauteous well: here’s our punctual for a beauteous location office:

You tin prime 1 of astir 2 twelve circumstantial styles to emulate, and pre-set the facet ratio of the photo. Here’s our New York java store with the Moody style applied:

Here, we statesman to spot Magic Media’s biggest weakness creeping in: uncanny vale photorealism.

Here’s different banal photograph effort that almost looks good… but for the deformed hands, confusing limb physics, and inheritance ensemble of melty-faced monsters:

It’s utile for generating vector creation too, and the images tin beryllium exported straight arsenic PNGs with nary background, but the images themselves are a small amateurish.

Here’s our key-holding cartoon fig again, this clip holding a perfectly creaseless cardinal successful 1 manus and a smaller, seemingly melted cardinal successful the other:

Here’s the terrifying effect of utilizing the aforesaid punctual with the 3D Chrome benignant applied:

Because Magic Media is embedded successful Canva, it’s incredibly casual to adhd text, resize the finished image, oregon adhd effects to the generated images. That’s a large plus, but successful my opinion, not capable to compensate for amateurish prime of the representation generation.

Here’s an illustration of however accelerated AI tools are developing. As I was penning this blog post, Google added AI representation procreation capabilities straight into Google Docs. Now, you tin usage the @image bid and select “Help maine make an image.”

It’s beauteous simple. You tin usage 1 of 3 facet ratios and specify 1 of six pre-determined styles, and Google returns 4 images to take from.

Here’s a decent small representation for the punctual “A cartoon quality with ginger hairsbreadth carrying a immense magnifying glass”:

And here’s “A cartoon quality with ginger hairsbreadth carrying a immense aureate key” with the Watercolor benignant applied:

Although these cartoons are decent, Gemini seems to person a peculiar skill: photography. It rendered beauteous scenes for my location bureau punctual with the Photography benignant selected:

And Gemini for Workspace seems to grip photos of radical adjacent better. Here’s a very realistic rendition of “Stock photograph of idiosyncratic moving connected their laptop successful a New York java shop”—even down to the Apple logo connected the laptop:

And here’s “Photo of a pistillate giving a speech connected stage”. I cannot archer this representation was AI-generated:

These images are tiny and low-resolution, but arsenic a large plus, you tin make them successful the travel of work—pretty utile for adding successful a speedy mock-up oregon placeholder to walk connected to your plan squad oregon amended successful the future.

This is evidently a precise caller diagnostic (when I tested it, representation procreation failed for maine astir 70% of the time), but I’d expect it to amended beauteous rapidly and go a large contender for champion AI representation generator.

Final thoughts

AI text-to-image generators are astatine their champion erstwhile you inquire for straightforward designs and don’t person a peculiarly beardown sentiment of the nonstop representation you privation to see. If you privation a speedy banal photograph oregon blog illustration, and don’t person to interest astir pesky marque guidelines, astir of these tools are up to the task (apart from possibly ChatGPT… sorry).

But the much circumstantial item you privation from the image—words, numbers, peculiar marque guidelines—and the stronger your sentiment astir what you privation the last representation to look like, the much frustrating these tools become.

I deliberation Adobe Firefly is the champion AI representation generator due to the fact that it sits astatine the intersection betwixt generative AI and accepted plan tools. It pairs each the originative benefits of AI with the editing power of Photoshop oregon Illustrator. That means it tin tackle analyzable plan workflows, similar creating a bid of cohesive characters, oregon applying peculiar styles oregon compositions. If you’re superior astir utilizing AI representation generators for your marque oregon business, I’d commencement with Firefly.

I’ll support updating this station arsenic caller AI representation generators are released and existing tools proceed to get updated. Want to inquire maine to reappraisal a instrumentality for you? Let maine cognize connected LinkedIn.