Adobe just made the most significant move in its ongoing campaign to own the enterprise AI image generation market. On Tuesday, the company launched a public beta of Firefly Custom Models — a feature that lets businesses train bespoke AI image generators on their own visual assets, producing outputs that match specific brand aesthetics, product designs, and stylistic guidelines. It’s a direct answer to a question that has dogged every major AI image tool: How do you make generic models useful for companies that need precise, on-brand creative output?
The answer, Adobe believes, is customization at scale.
As reported by The Verge, the new capability allows users to upload between 30 and 10,000 reference images — photos, illustrations, design assets — and use them to fine-tune Firefly’s underlying image generation model. The result is a custom AI that doesn’t just produce images in a general style but can learn the visual DNA of a particular brand, product line, or creative direction. Think a beverage company generating on-brand marketing visuals without starting from scratch every time, or a fashion label producing consistent lookbook imagery across dozens of campaigns.
Alexandru Costin, Adobe’s vice president of generative AI, told The Verge that the feature is designed to solve a fundamental limitation of general-purpose image generators. “The base model knows a lot of things, but it doesn’t know your things,” Costin said. That framing is telling. Adobe isn’t positioning Firefly Custom Models as a novelty or creative toy. It’s positioning it as infrastructure — the kind of tool that becomes embedded in a company’s creative workflow and, once adopted, is very difficult to rip out.
The Technical Architecture — and the Competitive Stakes
Custom Models builds on Adobe’s existing Firefly platform, which the company has carefully differentiated from competitors like Midjourney, Stability AI, and OpenAI’s DALL-E by training exclusively on licensed content and Adobe Stock imagery. That intellectual property firewall has been Adobe’s primary selling point to risk-averse enterprise customers worried about copyright exposure. Custom Models extends this logic: because businesses are training on their own proprietary assets, the legal and ethical ground is even firmer.
The technical process works through what Adobe calls Style and Subject training. Style training teaches the model to replicate a particular visual aesthetic — color palettes, lighting approaches, compositional tendencies. Subject training teaches it to recognize and reproduce specific objects or products. A shoe company, for instance, could train a model to generate images of a particular sneaker in various settings and contexts, maintaining accurate proportions, colorways, and design details across every output.
Training times vary. Adobe says most custom models can be ready in under an hour, depending on the volume of reference images and the complexity of the training task. That’s fast enough to be practical for iterative creative workflows — a designer could train a model in the morning and be generating usable assets by lunch.
But speed isn’t the whole story. The quality bar matters enormously here. Enterprise creative teams won’t tolerate outputs that are 80% right. A product shot where the logo is slightly wrong or the brand color is half a shade off is worse than useless — it’s a liability. Adobe says it has invested heavily in fidelity, particularly for subject training, where accuracy of physical details is non-negotiable. Whether the outputs consistently meet that bar in real-world production environments remains to be seen as the beta scales.
The pricing model also signals Adobe’s ambitions. Custom Models is available to subscribers of Adobe’s Firefly premium plans, which start at $9.99 per month for individuals and scale up significantly for enterprise accounts. Each custom model training run consumes generative credits, Adobe’s internal currency for AI operations. The company hasn’t disclosed exactly how many credits a typical training run costs, but the credit-based system ensures that heavy users pay proportionally more — a structure that could generate substantial recurring revenue if adoption takes hold across large organizations with dozens or hundreds of brands and sub-brands to manage.
Adobe isn’t operating in a vacuum. Competitors have been circling this same opportunity. Stability AI offers fine-tuning capabilities through its API. Midjourney has hinted at enterprise features. And a growing number of startups — companies like Astria, Scenario, and Bria — have built entire businesses around custom model training for specific verticals like gaming, e-commerce, and advertising. What Adobe has that none of these players can easily replicate is distribution. Firefly is integrated directly into Photoshop, Illustrator, and Adobe Express, tools that already sit at the center of most professional creative workflows. A custom model trained in Firefly can feed directly into the applications where designers actually work. No API calls. No exporting and importing between platforms. That integration advantage is enormous.
There’s also the data governance angle. Large enterprises — particularly those in regulated industries like financial services, healthcare, and consumer packaged goods — need assurances about where their training data goes, who can access it, and how it’s stored. Adobe says custom model training data is siloed per customer and not used to train Adobe’s general models. For a Fortune 500 brand uploading thousands of proprietary product images, that guarantee matters. It’s table stakes, really, but it’s table stakes that many smaller AI startups can’t credibly offer.
The timing of this launch is no accident. Adobe’s stock has been under pressure as investors question whether the company can translate its AI investments into meaningful revenue growth. The company reported strong fiscal Q1 2025 results in March, with total revenue of $5.71 billion, but guidance disappointed Wall Street, and shares sold off. Analysts have been looking for concrete evidence that Firefly is becoming a revenue driver rather than just a feature enhancement for existing subscriptions. Custom Models, with its credit-consumption model and enterprise pricing tiers, is Adobe’s clearest attempt yet to show that generative AI can be a standalone profit center.
Early enterprise adopters appear to validate the approach. Adobe has cited companies like Nickelodeon and IPG’s McCann agency as participants in earlier closed betas of Custom Models. The use cases range from generating consistent character imagery for entertainment properties to producing localized marketing assets at scale — work that traditionally required extensive photo shoots or hours of manual design iteration.
And yet, questions remain. Fine-tuning AI models on small datasets — even datasets of a few hundred images — can produce inconsistent results. Overfitting is a persistent risk: a model trained too narrowly might generate images that look eerily similar to the training data rather than producing genuinely novel variations. Adobe says it has built safeguards against this, but the public beta will be the real test. Thousands of users with wildly different training sets, quality expectations, and use cases will stress the system in ways internal testing cannot.
There’s a broader strategic dimension too. By encouraging businesses to invest time and resources in training custom Firefly models, Adobe is building switching costs. A company that has trained 50 custom models, integrated them into team workflows, and built creative processes around Firefly’s outputs isn’t going to casually migrate to a competitor. This is the classic enterprise software playbook — make the product indispensable not just through features but through accumulated investment. Adobe has executed this strategy before with Creative Cloud. It’s now attempting to replicate it with AI.
So where does this leave the market? The generative AI image space is fragmenting into two distinct tiers. At the consumer and prosumer level, tools like Midjourney and ChatGPT’s image generation compete on general quality, ease of use, and viral appeal. At the enterprise level, the competition is increasingly about customization, governance, integration, and intellectual property safety. Adobe is betting — with considerable justification — that it can dominate the second tier.
Whether that bet pays off depends on execution. The technology has to work reliably at scale. The outputs have to be good enough that creative directors trust them. And the pricing has to make economic sense compared to traditional production methods. Early signals are promising. But promising and proven are very different things, and Adobe’s investors are watching closely.
WebProNews is an iEntry Publication