Aiodic

Back to the table

94 Creative tool Ff Adobe Firefly Creative Tools

Adobe Firefly

Adobe's generative AI.

What it is

Adobe Firefly is the name Adobe gives to its family of generative AI models, and to the app where you can use them. Generative AI means software that creates new material, such as a picture, a video clip or a sound, from a written description. You type "a lighthouse on a cliff at dusk, painted in watercolour" and a few seconds later you have several versions to choose from.

What sets Firefly apart from most rivals is where it lives. Adobe makes Photoshop, Illustrator, Premiere Pro and Adobe Express, some of the most widely used creative programs in the world. Firefly is built into those programs, so many people use it without ever visiting a separate AI website. If you have clicked "Generative Fill" in Photoshop to remove a stranger from a holiday photo, you have used Firefly.

Here is what the Firefly family does, in plain terms. Feature names change often, but the main abilities have stayed much the same:

Adobe announced Firefly on 21 March 2023 as a beta, which means an early test version. From the start, Adobe's main selling point was not that Firefly made the most striking pictures. It was that Firefly was, in Adobe's words, designed to generate content "safe for commercial use". Adobe said its first model was trained on Adobe Stock images, openly licensed content and public domain content where copyright has expired. Adobe Stock is Adobe's own library of photos and illustrations that businesses pay to use, supplied by photographers and artists who earn a share of each sale.

That promise, usually shortened to "commercially safe", is the thread running through this whole article. It is a real difference from tools such as Midjourney and Stable Diffusion, which learned from images collected across the internet. But it has limits, and it has been tested more than once.

Under the hood, Firefly's image models belong to the same broad family as other modern image generators. The Diffusion Model page explains how that kind of system turns random noise into a picture, step by step. Adobe has not published full technical details of its models.

Why it matters

It brought AI image tools to people who never went looking for them. Millions of designers, photographers and small business owners already paid for Photoshop or Adobe Express. When Generative Fill appeared inside Photoshop, AI editing became one more button in a familiar program, not a separate website with its own habits to learn. Adobe said in October 2024 that Firefly had been used to generate more than 13 billion images since launch, and in February 2025 it put the figure at over 18 billion assets. Those are Adobe's own counts, but they show how quickly the tools became routine.

It made training data a selling point. Before Firefly, few people outside the tech world asked what an image generator had learned from. Adobe turned that question into marketing. By training on material it had a licence to use, and by offering businesses a legal promise to back that up, Adobe pushed the whole industry to answer questions about where training images come from and whether their creators were paid. You can read more about why training data matters on the Datasets page.

It shows how "safe" is a matter of degree. For a solo creator, the practical question is simple: if I put this image on a book cover or an advert, could it get me into trouble? Firefly's story is a useful case study because Adobe made strong claims, then faced reporting and customer anger that tested them. Knowing where the claim holds up, and where it doesn't, helps you judge every other tool's promises too.

It is becoming a hub for other companies' models. Since 2025 the Firefly app has also offered image and video models made by other companies, including OpenAI, Google and Black Forest Labs. That makes Firefly more useful, but it also means that not everything you make inside Firefly comes with Adobe's "commercially safe" promise. More on that below.

Real-world examples

The launch and the legal promise (2023)

When Adobe unveiled Firefly on 21 March 2023, it paired the training-data claim with two other commitments. It introduced a "Do Not Train" tag that creators could attach to their work to signal they did not want it used for AI training. And it said it was developing a way to pay Adobe Stock contributors whose images had helped train the model.

On 23 May 2023, Generative Fill arrived in the beta version of Photoshop. It let users add, extend or remove content in an image using a text prompt, and it became one of the most talked-about AI features of the year. On 8 June 2023, Adobe announced Firefly for businesses and said enterprises could obtain an IP indemnity for content made with certain Firefly-powered workflows. An indemnity is a contractual promise to cover the customer's legal costs if someone sues them. In this case, it covered claims that a Firefly image infringed someone's copyright or other rights. Speaking to TechCrunch, Adobe's general counsel made clear it covered the Firefly output itself, not things a customer added to it afterwards, such as a famous cartoon character.

On 13 September 2023, Firefly came out of beta. Generative Fill and Generative Expand became part of the standard version of Photoshop, and Firefly was also built into Illustrator and Adobe Express. Adobe introduced "generative credits", a monthly allowance of AI actions included in each plan. At launch, one credit covered one Generative Fill or Expand on an image up to 2,000 by 2,000 pixels, and people who ran out could keep going at slower speeds. The same day, Adobe paid its first Firefly bonus to eligible Adobe Stock contributors. Adobe did not disclose the amounts. According to Adobe, the bonus was based on how many of a contributor's images were in the training data and how often their work was licensed, and it said the payment would continue yearly.

Two limits are worth noticing. First, the indemnity was for business customers on enterprise contracts, not for individuals on an ordinary subscription. Second, the bonus was paid at Adobe's discretion and varied from person to person. Contributors had not negotiated it, and many had uploaded their work under Adobe Stock's terms long before AI training was on anyone's mind.

The Midjourney pictures inside Firefly (April 2024)

In December 2022, a few months before Firefly launched, Adobe Stock began accepting pictures made with generative AI tools, provided contributors labelled them as such. Some of those submissions were made with rival generators such as Midjourney.

On 12 April 2024, Bloomberg reported that some of these AI-made Stock images had ended up in Firefly's training data. According to Bloomberg, AI-generated images made up about 5 percent of the material used to train Firefly, and Adobe had not mentioned this in its public presentations about why Firefly was safer than the competition. The awkward part was obvious: Midjourney had learned from images collected across the internet, which is exactly the practice Adobe had positioned Firefly against.

Adobe did not dispute that AI images were in the mix. Its response, as reported by Bloomberg, was that every image submitted to Adobe Stock goes through a moderation process meant to screen out trademarks, recognisable characters, logos and references to artists' names, and that AI images were only a very small part of the training set.

Does this make Firefly unsafe? Not necessarily. A small share of second-hand AI images, filtered by human moderators, is a very different thing from scraping the whole web. But it showed that "trained only on licensed content" is harder to guarantee than it sounds, because a licence tells you who uploaded an image, not how that image was made.

The Terms of Use backlash (June 2024)

In early June 2024, Photoshop users were asked to accept updated Terms of Use before they could keep working. Screenshots spread quickly on social media. The wording said Adobe might access user content through both automated and manual methods, and many artists, including some working under confidentiality agreements with clients, read it as Adobe giving itself the right to look at their unpublished work, and perhaps to train Firefly on it.

On 6 June 2024, Adobe published a clarification. It stated that Adobe does not train Firefly on customer content and will never assume ownership of a customer's work, and explained that the changes were about how it reviews content to screen for illegal or abusive material. Later that month Adobe rewrote the terms with plain-language summaries, saying it does not scan content stored only on your own device and does not train generative AI on customer content, except for material submitted to Adobe Stock. The episode mattered because Firefly's whole reputation rested on trust about data. Even a badly worded legal update could shake it. For more on what companies can do with the files you upload, see the Privacy page.

Video, and then other companies' models (2024 to 2025)

On 14 October 2024, Adobe released a limited beta of the Firefly Video Model, including Generative Extend in Premiere Pro, which adds a few extra frames to the end of a clip that is slightly too short. Adobe described it as the first publicly available video model designed to be commercially safe. On 12 February 2025, the video model moved into a wider public beta inside a new Firefly app, with new paid Firefly plans.

Then Adobe changed course. On 24 April 2025, at its MAX London event, it announced that the Firefly app would also offer outside "partner models", starting with image and video models from Google and OpenAI and Black Forest Labs' Flux 1.1 Pro, with more promised later. In September 2025, partner models arrived in Photoshop's beta Generative Fill as well, as the Flux page describes. At the time of writing, Adobe's partner page also lists models from Runway, Kling, Luma AI and ElevenLabs, among others.

This is where the commercially safe promise needs careful reading. Adobe describes its own Firefly models as commercially safe. It does not make that claim for the partner models, which were trained by other companies on data Adobe did not choose. Its partner page says it is the creator's responsibility to decide whether a partner model is appropriate for a project, including whether it is safe for commercial use. Adobe does attach Content Credentials to images made in the Firefly app, recording whether a Firefly model or a partner model was used. Content Credentials are a kind of digital label stored inside the file, based on the C2PA standard explained on the OpenAI Images page. Adobe co-founded the Content Authenticity Initiative, the group that promotes these labels, with The New York Times Company and Twitter in 2019.

Adobe has kept adding to Firefly since, including Firefly Image Model 5 and soundtrack and voice-over tools in October 2025, and a conversational assistant that can carry out tasks across Adobe's apps, announced in April 2026. The details change quickly, so check Adobe's own announcements for the current picture.

An everyday scenario: extending a cover for a paperback

This is an illustration, not a real case. Priya has licensed a photo of a stormy harbour for her self-published mystery novel. It works perfectly as an ebook cover, but the paperback needs a wrap-around cover with a spine and back panel, and the photo is too narrow. In Photoshop she enlarges the canvas and uses Generative Expand to continue the sea and sky to the left. Two of the three results show the harbour wall bending at an odd angle; the third is convincing. She then uses Generative Fill to remove a modern crane from the skyline that clashes with the book's period setting. She checks the result at full size for smudges and repeated patterns, keeps a note of which parts are AI-made, and reads Amazon KDP's rules on AI content before uploading. Finally, she checks the licence for the original photo, because extending a picture doesn't change what its licence allows.

Try it yourself

Menus, plans and credit rules change often, so treat these steps as a guide to the process. At the time of writing, Adobe offers a free plan with a limited number of generations, and some features and models need a paid plan. Check the current details on Adobe's Firefly plans page before you rely on any of this.

  1. Sign in at firefly.adobe.com. You need a free Adobe account. Look at how many free generations or credits your account shows, so you know how many experiments you can run.
  2. Check which model is selected. Before you type anything, find the model picker. Note whether it is set to an Adobe Firefly model or a partner model from another company. For work you may sell, start with an Adobe Firefly model, because that is the one Adobe calls commercially safe.
  3. Generate a simple image. Try a prompt like this one, which you can copy: A cosy independent bookshop at night, warm light spilling onto a rainy street, no people, photographic style Look at all the results and pick your favourite.
  4. Change one thing at a time. Add wide shot, low angle or swap photographic style for watercolour illustration. Changing one detail per attempt teaches you what each word does. The Prompt page has more tips.
  5. Try Generative Fill on your own photo. Upload a photo you took, such as a desk or a garden. Paint over an object you want gone and leave the prompt empty to remove it. Then paint over an empty area and type something to add, like a small potted plant.
  6. Try expanding the frame. Use the expand option to turn a square photo into a wide banner. Look closely at the new edges. This is where AI images most often go wrong, with warped lines, repeated patterns or objects that make no sense.
  7. Compare with a partner model. If your plan allows it, run the same prompt through a partner model and compare the results. Notice that the quality may be different, and so are the promises attached to it.
  8. Check the label. Download one image and upload it to the Content Credentials inspection tool (verify.contentauthenticity.org). See what it records about how the image was made.
  9. Make a social graphic in Adobe Express. Adobe Express, Adobe's simpler design app, has Firefly built in and a free plan of its own. Try making a square post announcing a book or product, using an AI image as the background and your own text on top.

Common mistakes and misunderstandings

Assuming "commercially safe" means "no risk at all". Adobe's claim is about how its own models were trained. It does not mean an image can never resemble someone else's work, and it doesn't protect you from things you add yourself, such as a brand logo or a famous character. The legal indemnity Adobe advertises is for business customers on enterprise contracts, with conditions. If you are an individual on an ordinary plan, read the current terms rather than assuming you are covered.

Forgetting that partner models are different. Inside the same Firefly app you can now pick models from several companies. Adobe's commercially safe description applies to its own Firefly models, not to the partner models. Check the model picker every time, especially if a picture is going on something you sell.

Thinking you own the copyright in an AI image. Using Firefly does not by itself make an image your copyrighted work. In the United States, the Copyright Office has said that material generated by AI from a prompt is not protected on its own; the Midjourney page covers the key rulings. Your own edits and arrangement may be protected, but others may be free to copy the purely AI-made parts.

Believing Adobe trains Firefly on your files. Adobe has stated publicly that it does not train Firefly on customer content. Anything you submit to Adobe Stock for sale is a different matter, because Stock content is part of the training library. If you contribute to Stock, read the contributor terms.

Running out of credits halfway through a project. Generative credits are the fuel for most Firefly features, and different features and models cost different amounts. Video usually costs far more than images. At the time of writing, standalone Firefly plans start at around 10 US dollars a month, and Creative Cloud plans include their own allowance. Plans and prices change often, so check before starting a big job.

Ignoring the platform's disclosure rules. Amazon KDP, for example, asks authors to disclose AI-generated images, including covers, while content only edited or improved with AI tools counts as AI-assisted and doesn't need disclosing. Where an extended or partly generated cover falls is a judgement call, so read the current rules, and when in doubt, disclose.

Removing the labels. Adobe's terms forbid removing or altering watermarks or Content Credentials to mislead people about where an image came from. Free-plan outputs have at times carried a visible watermark. Don't crop it out; upgrade or use a different image instead.

Skipping the final check. AI fills can look perfect at thumbnail size and fall apart when printed. Zoom in on hands, text, edges and anything that should be a straight line. The AI doesn't know your book is set in 1920, so it may add a modern car in the background.

Related elements