Adobe is the undisputed king of photo and design software, and that alone makes its Firefly AI image generator worth paying attention to. Firefly has some nifty capabilities for creating art or faux photography to your taste, packaged in a typically excellent Adobe interface. It also powers features within some of the company’s other apps—Photoshop and Illustrator in particular—and it integrates with Adobe Express, the company’s lightweight online design tool. Despite all that, Firefly’s image results are disappointing, just as they are with other AI image generators. How Do You Get Firefly?You can try out Firefly in its purest form on its own website, firefly.adobe.com. There, you can see all the shiniest, newest capabilities. You can also access some Firefly features in Photoshop, Express, Adobe Stock, InDesign, and Illustrator. For example, in Photoshop, you can use Firefly to replace the background of an image with an entirely new one that you describe in words. This review focuses on the standalone version of Firefly on the web.
Since 1982, PCMag has tested and rated thousands of products to help you make better buying decisions. See how we test.
Since generative AI is such an intense computational activity that requires processing on Adobe’s servers, you don’t get unlimited access to it with your Creative Cloud subscription. Instead, your subscription comes with Generative Credits. Most generating actions use up one credit for standard images of up to 2,000 by 2,000 pixels. By comparison, Microsoft gives you Boosts for Copilot, which speeds up image generation; after you run out of Boosts, your images come slowly.Spec-wise, you need ChomeOS, macOS 12 or later, Windows 10, or Windows 11 to run Firefly. Adobe doesn’t mention any CPU or graphics requirements, only that you have at least 4GB of main RAM. You can also use the web app on Android 9 or later and iOS 15 or later.
How Much Does Firefly Cost?You can use the Firefly web app for free, but you have to sign in with a Creative Cloud, Adobe Firefly, or Adobe Express account—it works even if you have a free plan. With this free access, you get 25 image credits, and your results are watermarked. At least the interface doesn’t show ads. A $4.99-per-month (or $49.99-per-year) Firefly Premium plan increases your image credits to 100, removes watermarks, and gives you 100GB of online storage for your images. Beyond that is the $9.99-per-month (or $99.99-per-year) Adobe Express Premium plan, which gets you 250 credits and all the Express Premium features, like templates and animation effects.Firefly charges the lowest monthly rate for a premium account among similar services. For comparison, ChatGPT doesn’t have image generation in its free account; it requires a $20-per-month Plus account for that. Midjourney doesn’t have a free plan at all, and its Basic plan ($10 per month or $96 per year) only lets you have three concurrent jobs in progress and 10 jobs waiting in the queue at any given time, plus other limitations that limit how quickly the AI tool will make your artwork. OpenArt gives you 50 image credits for free and charges $12 a month or $72 a year for its Starter plan. Stable Diffusion has a free plan with ads, watermarks, and non-priority processing (read: slow), or you can pay $84 per year for a Pro account, which removes the ads and watermarks and gives you 2,000 fast image generations. Meta AI generates images for free if you log into a Facebook account. Which of these is the best value? It depends on what features you want, but Firefly’s combination of free functionality and low cost for Premium are appealing, value-wise.What Can You Do With the Firefly Web App?Firefly on the web can generate images based on text prompts you enter, which can describe any outlandish visual idea you get. For example, you can tell it to create a realistic photo of a reindeer on a Caribbean beach with owls flying around and a team of huskies pulling a dogsled with a driver on the shore, as I did to see what Firefly can do.
(Credit: Adobe/PCMag)
The way you word your text prompt is critical to how much the image matches your expectations, and even when you think you have it perfect, you sometimes see anomalies and distortions. For example, I specifically asked for a team of husky dogs pulling a dogsled, but in all the results, there were no dogs. And notice that I asked for and didn’t get a dogsled driver and did not ask for (but got) beach loungers.The prompt I wrote would create a clear image in most people’s minds—including a dogsled with a musher—but Firefly and most other image generators could not come up with the expected visual. ChatGPT/Dall-E put an owl in the driver’s seat after I specifically told it to make the driver human, and it even replied that it had done so. Stable Diffusion didn’t include a dogsled at all and put spruce trees in the Caribbean landscape I’d specified. You could say that Stable Diffusion was running with my idea of putting cold-weather things into the tropics, but Firefly kept the setting true to the Caribbean one I’d specified. OpenArt.ai came close, except the dogsled passengers were facing backward, and it planted evergreens in the tropics, too. Note that OpenArt lets you choose among various AI models, including its own OpenArt SDXL and Stable Diffusion—80 models in all.Constraints of the Prompt BoxMidjourney and Stable Diffusion helpfully include a negative prompt box to rule out stuff you don’t want to see in your image. Firefly doesn’t, and when I tried giving it a negative instruction, “no clouds,” I still got plenty of clouds. Putting “clouds” in Stable Diffusion’s Negative prompt box did the trick. OpenArt.ai has a negative text box, too, but perhaps even more interesting is its Prompt Adherence slider, which determines how strictly the model interprets your text, as opposed to taking liberties. Firefly doesn’t have anything like it. In practice, with the adherence slider cranked up, however, I got weird cartoonish results, even when my prompt asked for photorealism.After creating your image, Firefly leaves your text in the prompt box. You can only add to it if you want to tweak the same scene. You can’t continue the conversation for these edits as you can with Copilot, ChatGPT, Gemini, and Meta AI.Additional Controls for Making ArtworkFirefly has some controls that you don’t find in all AI image generators. Chief among them is the ability to select an aspect ratio. Firefly is better than Gemini, Meta AI, and Copilot because you can choose to generate widescreen or vertical images, whereas the others only produce squares. Firefly has a setting for Aspect Ratio in the left panel, with choices of Landscape (4:3), Portrait (3:4), Square, and Widescreen (16:9). These choices are more useful than Stable Diffusion’s aspect ratio choices, which don’t include those standard dimensions. Of the services I tested, ChatGPT Pro with Dall-E 3, OpenArt.ai, Stable Diffusion, and Midjourney let you specify different aspect ratios, including widescreen. ChatGPT/Dall-E lets you change the aspect ratio right in the chat prompt, and OpenArt has all the standard aspect ratios.
(Credit: Adobe/PCMag)
You can also choose to use either version 2 or 3 of Firefly. Version 3 is marked as being in Preview, but it worked fine in testing, delivering more convincing images with more detail and realism.When making images in Firefly, you also get to decide whether you want Photo or Art style and apply various effects to your image. You can choose from a good selection of effects, such as Bokeh, Pointillism, Hyper Realistic, and Antique Photo. If you choose the Photo style, you get more settings specific to that image type, including Aperture, Shutter Speed, and Field of view (wide-angle or telephoto, for example). In the Art style, a good selection of mediums and textures are available in categories such as Acrylic and Oil, Watercolor, Pencil, Architectural sketch, 3D, Dramatic lighting, and Neon, to name just a few. But when I asked for an image in the style of Rembrandt, I got the images below, which hardly do justice to the Dutch master.
(Credit: Adobe/PCMag)
Firefly has a Suggestions slider that tries to fill out a prompt if you just enter a word or two. For example, I turned on the slider and wrote “sun,” and the more creative suggestions below popped up above the prompt box. Suggestions could be useful when you need ideas for your image.
(Credit: Adobe/PCMag)
This feature is somewhat similar to OpenArt.ai’s Enhance Prompt feature, which tries to improve your prompt using more AI, but that works with longer and more detailed prompts you enter as well.Working With Uploaded ImagesFirefly can’t create an image based only on another that you upload along with a prompt, nor can it modify an uploaded image for you. It does, however, let you upload an image to use as a reference to base the style or the “structure” of the result on. You can’t blend two images as you can with Midjourney, though. Stable Diffusion has three self-explanatory options for using your own content: Copy Contour, Copy Pose, and Sketch to Image. OpenArt gives you several more ways to influence your result image based on an uploaded picture. In this case, Firefly is more limited than those other services.Remember my not-so-successful attempt at creating an image in the style of Rembrandt? After I uploaded an actual Rembrandt self-portrait, I got the improved result below.
(Credit: Adobe/PCMag)
It’s still not going to convince art historians, but it’s much closer. For comparison, here are side-by-side results using the same uploaded Rembrandt picture and prompt text from Firefly, ChatGPT 4, Stable Diffusion, and OpenArt.ai.
(Credit: Adobe/OpenAI/Stability AI/OpenArt/PCMag)
Stable Diffusion performed better than Firefly and ChatGPT with Dall-E 3, but OpenArt performs this task best for my money, using its SDXL model. Meta AI doesn’t let you upload sample images, and remember that Gemini won’t do anything involving images of people at the moment.Dismal Text GenerationYou can tell Firefly to add specific text to your image in your prompt, but you may end up with a garbled mess like I did.
(Credit: Adobe/PCMag)
I asked it to add “cheers to another great campaign.” Doing this in ChatGPT, Meta AI, and Gemini yielded the correct text. Midjourney also generates text you put in quotation marks, but OpenArt.ai is not interested in text at all, doing nothing when I told it to add some. In its favor, Adobe does point you to Adobe Express for adding text, a tool better suited to that task, but it would do well to emulate OpenArt and simply ignore text prompts asking it to include titles or captions. Serious designers want tools meant specifically for text when adding text rather than AI generation, at least for now. Someday, it could become something that AI will do with aplomb.Editing and Sharing Your Firefly CreationsAfter Firefly generates your image, you have a few choices for the next steps. You can use more Adobe AI tools like Generative Fill. You can have it generate another similar set of images. You can also use the structure of the generated image as the basis for another new image. Surprisingly, Copilot does a more Adobe-ish thing by highlighting objects in your created image as you hover the mouse over them and letting you blur them or punch up their color. With Firefly, you can also send the image to Adobe Express to turn your art into a social post, complete with text and additional graphics.
(Credit: Adobe/PCMag)
When you download or share an image created by Firefly, Adobe applies Content Credentials metadata to indicate the work was created by AI. If your request is based on a sample you upload, the metadata indicates that as well. Adobe has long been pushing for this kind of transparency and has even established a Content Authenticity Initiative.What Won’t Firefly Do?Unlike Gemini, which currently won’t produce any images that include humans due to detected bias in its algorithm, Firefly can produce presentable people in your results. You won’t, however, see anything higher than G-rated, so don’t hope to create adult images. Don’t expect to include anything trademarked, either. When I asked for a group of friends drinking Coca-Cola, the resulting beverages looked more like white wine than Coke. That’s because Adobe prevents copyright infringement in Firefly. The service also has, as you might expect, a policy against creating any images that are “harmful or offensive content that could promote violence, hate speech, or other harmful activities,” including misinformation and disinformation. If you cross any of these lines (Adobe uses both machine and human reviewers to detect them), Adobe may disable your Firefly account.Does Adobe Train Its AI With Your Content?According to Adobe’s documentation, “We don’t train on any Creative Cloud subscribers’ personal content.”That said, the company is aware that some users want to train an AI with their images to, for example, match a corporate style, and Adobe says it is looking into giving those users something. Adobe does train on images in Adobe Stock, the company’s answer to services like Shutterstock and Getty Images. This policy is in stark contrast with Meta AI and Google Gemini, which take advantage of their massive social networks (Facebook, Instagram, Google Photos, and YouTube) for training fodder. Both claim to train using only public images and videos, though.Should You Fly With Firefly?Anyone interested in visual arts should pay attention to the emerging world of generative image AI. However, even the latest Firefly model is not going to put many graphic designers or photographers out of work, though it does suffice for creating visuals that are either totally generic or absurdly whimsical. It’s the images between those two extremes, however, that are most useful in the real world. The field of generative AI art tools is so new that we haven’t evaluated enough of them to select an Editors’ Choice winner, but Adobe Firefly is a contender, along with Dall-E, OpenArt.ai, and Stable Diffusion. Your choice depends on how much control you need and whether you prefer a conversational interface. At this point, OpenArt appears to be better than the others in terms of AI model choice, controls, and usability.
Pros
Good selection of preset styles and effects
Can match the style or structure of uploaded images
Choice of aspect ratios
Low monthly fee for Premium
Your content doesn’t train Adobe’s AI
View
More
Cons
Images can look distorted or not match the request
No negative prompt option to exclude objects from images
Requires subscription to remove watermark
The Bottom Line
Adobe’s Firefly is a promising entrant in the crowded field of image-generative AI with lots of style options, but it doesn’t produce completely convincing images.
Like What You’re Reading?
Sign up for Lab Report to get the latest reviews and top product advice delivered right to your inbox.
This newsletter may contain advertising, deals, or affiliate links. Subscribing to a newsletter indicates your consent to our Terms of Use and Privacy Policy. You may unsubscribe from the newsletters at any time.
