Adobe's Firefly Bet Isn't Working

In September 2023, Adobe CEO Shantanu Narayen stood on stage at Adobe MAX in Los Angeles and made a promise that would define the company's AI era. Firefly, Adobe's family of generative AI models, would be the "commercially safe" choice for creative professionals and enterprises — trained exclusively on licensed content, indemnified against IP claims, and integrated natively into the Creative Cloud tools that 35 million people already use.

"We believe creators should be at the center of AI," Narayen told the audience. "Not replaced by it. Not exploited by it. At the center."

The crowd applauded. The stock ticked up. And for a brief moment, it looked like Adobe had found the perfect positioning in a chaotic market: the responsible AI company, the grown-up in a room full of move-fast-and-scrape-everything startups.

Eighteen months later, that positioning is looking less like a moat and more like a trap.

Is Adobe Firefly Actually Good Enough?

The most uncomfortable question in Adobe's boardroom is one that no earnings call has directly addressed: is Firefly's output quality competitive with the tools professional creatives actually use?

The data suggests it is not.

In the February 2026 Artificial Analysis Image Arena, which aggregates blind human preference rankings across thousands of side-by-side comparisons, the results are stark:

Sixth place. Behind every major competitor except the open-source baseline. And the gap is not marginal — Firefly's ELO rating sits 107 points below Midjourney, a difference that in blind testing translates to users preferring the competitor's output roughly 65% of the time.

A January 2026 survey by Blind of 2,400 professional designers, illustrators, and creative directors found that only 18% used Firefly as their primary AI image generation tool. Midjourney led at 41%, followed by Stable Diffusion variants at 22% and DALL-E at 14%. The remaining 5% used Flux, Ideogram, or other tools.

"Firefly is fine for social media thumbnails and placeholder assets," one creative director at a Fortune 500 consumer brand told me, requesting anonymity because of an active Adobe enterprise agreement. "But anything that needs to look genuinely compelling — hero images, campaign visuals, concept art — we're in Midjourney. It's not even close."

Why Did Adobe Choose Legal Safety Over Output Quality?

The answer is structural, and it reveals a tension that may be irreconcilable.

Adobe's Firefly models are trained on three categories of data: Adobe Stock's library of approximately 400 million licensed images, openly licensed content from sources like Wikimedia Commons, and public domain works. This was a deliberate choice. While Midjourney, Stability AI, and OpenAI trained their models on LAION-5B and similar datasets scraped from the open internet — billions of images harvested without explicit creator consent — Adobe chose to use only content it had clear legal rights to.

The rationale was sound, and it was driven by two forces:

First, litigation risk. By early 2024, multiple class-action lawsuits had been filed against Stability AI, Midjourney, and DeviantArt, alleging copyright infringement in training data. Getty Images sued Stability AI in both US and UK courts. The legal landscape was, and remains, genuinely uncertain. Adobe's bet was that enterprises — its most lucrative customer segment — would pay a premium for IP-clean AI outputs.

Second, stock-photo partnerships. Adobe saw an opportunity to turn its Stock library and licensing relationships into a competitive advantage. It signed expanded agreements with Getty Images and Shutterstock, creating a contributor compensation fund that promised to pay photographers and illustrators when their work was used to train Firefly. The deals were structured as revenue shares, with contributors receiving payments based on the frequency with which their assets influenced model outputs — a metric that is, in practice, nearly impossible to calculate with precision.

David Wadhwani, Adobe's president of digital media, told Bloomberg in mid-2024 that the licensed-data approach was "not a constraint but a competitive advantage." He argued that enterprise buyers would ultimately choose the tool that eliminated legal risk, even if it meant accepting some quality trade-offs.

Eighteen months later, enterprise buyers have not shown up in the numbers Adobe projected.

How Bad Is the Enterprise Adoption Problem?

Adobe does not disclose Firefly-specific revenue. This is, in itself, revealing. The company breaks out Digital Media segment revenue ($13.1 billion in FY2025), Creative Cloud revenue (approximately $11.4 billion), and total generative AI credit consumption (over 16 billion cumulative credits used since Firefly's launch). But it does not say what Firefly contributes in actual dollars.

Analysts have tried to back into the number. Morgan Stanley's Keith Weiss estimated in a January 2026 note that Firefly generated approximately $500 million in annualized revenue — a combination of generative credit upsells within Creative Cloud ($4.99/month for additional credits), standalone Firefly subscriptions ($9.99/month), and API licensing deals. Bank of America's Brad Sills put the figure slightly lower, at $400-450 million.

Both estimates are well below the $1 billion annual run-rate that Adobe's leadership guided toward at its 2023 analyst day.

The gap matters because it undermines Adobe's entire narrative. If Firefly's commercially safe positioning was going to command premium pricing and drive Creative Cloud ARPU expansion, the revenue should be accelerating by now. Instead, the evidence suggests that:

• Free-tier usage is high, paid conversion is low. Adobe bundles 25 generative credits per month with every Creative Cloud subscription. The majority of users consume their free allocation and never upgrade. Adobe's disclosure of "16 billion cumulative credits used" sounds impressive until you divide it by 35 million Creative Cloud subscribers over 18 months — it averages roughly 25-30 credits per user per month, barely above the free allocation.

• Enterprise pilots are converting slowly. Several large enterprise customers I spoke with described a similar pattern: IT or brand teams evaluate Firefly, approve it for "low-risk" use cases (internal presentations, draft concepts, social media filler), but continue using Midjourney or Stable Diffusion for high-visibility creative work. The indemnification promise is valued in theory but has not changed actual procurement behavior at scale.

• API revenue is modest. Adobe's Firefly API, launched in mid-2024, competes with OpenAI's DALL-E API, Stability AI's API, and Midjourney's nascent API. Pricing is competitive ($0.04-0.08 per image depending on resolution and model version), but adoption among developers and SaaS platforms has been limited. Most app developers building AI image generation features default to open-source models (Flux, Stable Diffusion) that can run on their own infrastructure at near-zero marginal cost.

> "Adobe's pitch is: you're paying for safety. But our legal team reviewed the actual IP risk of using Midjourney for marketing assets and concluded it was low enough to accept. So we're paying less for better output." — VP of Marketing at a Fortune 200 consumer goods company

Are Adobe's Stock-Photo Partners Getting a Fair Deal?

The partnerships that were supposed to make Firefly's training data an asset are becoming a source of friction.

When Adobe announced its contributor compensation program in 2023, it was framed as a model for how AI companies should work with creators. Photographers and illustrators whose Adobe Stock submissions were used to train Firefly would receive an annual bonus payment from a dedicated fund. The fund was seeded at $25 million annually and was expected to grow proportionally with Firefly revenue.

Two years in, contributors say the payments are negligible.

According to interviews with six Adobe Stock contributors who participate in the Firefly bonus program, annual payments have ranged from $18 to $340, with the median around $75. For context, many of these contributors have portfolios of 5,000-20,000 images on Adobe Stock and generate $10,000-50,000 per year in traditional licensing revenue.

"I got a Firefly bonus of $62 last year," said one contributor with over 12,000 images in the Adobe Stock library. "I spent more on the electricity to edit and upload those photos than Adobe paid me for training their AI on them."

The math is not hard to check. If Adobe's contributor compensation fund is approximately $25-35 million annually and there are roughly 300,000 active Adobe Stock contributors, the average payout works out to $80-115 per contributor per year — before accounting for the fact that distributions are weighted toward high-volume contributors and popular content categories.

Getty Images and Shutterstock, meanwhile, are navigating their own discomfort. Both companies signed data licensing deals with Adobe, reportedly worth $50-100 million annually combined. But those deals were predicated on the assumption that Firefly would become the dominant enterprise AI image tool — driving new revenue that would offset the cannibalization of traditional stock photo licensing.

That cannibalization is happening. Traditional stock photo revenue is declining 15-20% year-over-year across the industry. But the Firefly revenue that was supposed to replace it has not materialized at the projected scale. Getty Images CEO Craig Peters acknowledged on a Q3 2025 earnings call that "the transition from traditional licensing to AI-enabled content creation is taking longer than anticipated," which is corporate-speak for "the checks are smaller than we expected."

The Shutterstock Renegotiation

Shutterstock's deal with Adobe is reportedly up for renegotiation in mid-2026. Multiple sources familiar with the discussions say Shutterstock is pushing for guaranteed minimum payments rather than revenue shares — a signal that the stock-photo company has lost confidence in Firefly's growth trajectory. Adobe is reportedly resisting, preferring to keep the economics variable.

If Shutterstock walks or extracts significantly better terms, it could increase Adobe's cost of training data at exactly the moment when competitors are training on exponentially larger datasets at lower marginal cost.

Is Adobe's Stock Price Reflecting the Firefly Problem?

Adobe's stock tells a story of declining confidence in the AI narrative.

After peaking at approximately $700 per share in late 2024 following the initial Firefly hype, Adobe shares have traded in a $450-550 range through early 2026 — a roughly 25-30% decline from the peak. The company's price-to-earnings ratio has compressed from approximately 45x to 32x, closer to legacy software companies like Oracle and SAP than to AI leaders like Nvidia or even Salesforce.

The bear case on Adobe — articulated by analysts at Bernstein and Piper Sandler — is that Firefly's underperformance is not a temporary gap that will close with better models. It is a structural consequence of a constrained training dataset that will always lag competitors with access to larger, more diverse data. Every six months that Firefly remains behind on quality, more creative professionals build workflows around other tools — workflows that are sticky and hard to reverse.

The bull case, advanced by Goldman Sachs and JPMorgan, is that the legal landscape will eventually vindicate Adobe's approach. If courts rule that training on copyrighted data without consent constitutes infringement — a plausible outcome given pending cases — Midjourney and Stability AI could face injunctions, damages, or forced model retraining. In that scenario, Adobe's clean-data advantage becomes decisive overnight.

The problem with the bull case is timing. The major AI copyright cases are not expected to reach final resolution before 2027 or 2028. By then, the market may have already decided.

What Should Adobe Do Now?

Adobe has three options, none of them comfortable.

Option 1: Double down on the current strategy. Continue improving Firefly within the licensed-data constraint, invest in model architecture to close the quality gap, and wait for the legal environment to shift in its favor. This is the current path. The risk is that the quality gap never fully closes and the legal shift never arrives — or arrives too late to matter.

Option 2: Expand the training dataset. Strike new licensing deals with additional content libraries, individual creators, and possibly even social media platforms to dramatically increase the volume and diversity of Firefly's training data. Adobe has reportedly had exploratory conversations with Pinterest and Tumblr about content licensing deals, though nothing has been announced. This approach could narrow the quality gap but would significantly increase training data costs at a time when competitors' marginal data costs are near zero.

Option 3: Acknowledge the gap and integrate competitors. Rather than trying to make Firefly the only AI image generation tool in Adobe's ecosystem, allow users to plug in Midjourney, DALL-E, or Flux models directly within Photoshop and Illustrator. Adobe already supports third-party plugins — extending this to AI model selection would concede that Firefly is not the best model while preserving Adobe's position as the essential creative workflow platform. This is the most strategically sound option but the hardest one politically, because it would effectively admit that the last two years of Firefly investment have not achieved their primary goal.

The Canva Pressure

Adding urgency to Adobe's decision is Canva's aggressive AI integration. Canva has taken a pragmatic approach to AI — using a combination of its own models, licensed Stable Diffusion variants, and third-party APIs to power its Magic Design suite. Canva does not make grand claims about training data ethics. It simply ships the best output it can, as fast as it can, to its 200 million users.

For the non-designer majority — the marketers, educators, and small business owners who represent the largest growth opportunity in visual content creation — Canva's "good enough AI with great UX" is more compelling than Adobe's "legally safe AI with professional UX." And Canva's $3.8 billion revenue run-rate, growing at 55% annually, suggests the market agrees.

The Deeper Problem: Has Adobe Misread What Creators Actually Want?

There is a more fundamental critique of Adobe's Firefly strategy that goes beyond model quality and training data. It is that Adobe built Firefly for the enterprise procurement officer, not for the creative professional.

The emphasis on IP indemnification, commercially safe training data, and enterprise compliance features assumes that the buyer of AI creative tools is a legal or IT department. But the actual users — the designers, illustrators, photographers, and art directors who choose which tools to open every morning — make decisions based on output quality, creative flexibility, and workflow speed.

Every creative professional I interviewed for this article said some version of the same thing: "I don't care about indemnification. I care about whether the image looks good."

This is the same mistake Microsoft made with Bing in the early search wars — building for the channel partner and enterprise IT buyer while Google built for the end user. It is the same mistake BlackBerry made by optimizing for corporate security while iPhone optimized for user experience. The enterprise buyer eventually follows the user, not the other way around.

Shantanu Narayen has led Adobe through multiple successful transitions — from boxed software to subscriptions, from desktop to cloud, from creative tools to marketing automation. Each transition required the company to cannibalize existing revenue streams in pursuit of larger ones. The question now is whether Narayen and Wadhwani are willing to do it again: to acknowledge that Firefly's legal-safety-first approach has produced a product that is not competitive, and to take the painful steps necessary to close the gap.

The 16 billion Firefly credits consumed to date prove there is demand. The sixth-place quality ranking proves the product is not meeting it. And the $400-600 million in estimated revenue — in a generative AI image market that Goldman Sachs projects will reach $15 billion by 2028 — proves the window is closing.

Adobe still has the distribution, the brand, the enterprise relationships, and the creative workflow dominance to win this market. But winning requires building the best product, not just the safest one. And right now, Firefly is optimized for a courtroom that may never convene, while its competitors are optimized for the studio where creative work actually happens.