Adobe is not a startup. It’s a 40-year-old company that employs over 29,000 people and generates $19 billion in annual revenue. Its products — Photoshop, Illustrator, Premiere, Acrobat — are used by most professional creatives on the planet. When AI image generation emerged in 2022, Adobe faced a decision that was fundamentally different from the one facing OpenAI or Stability AI: how do you introduce AI to an existing professional user base that has built its career on the software you make?

The creative community’s reaction to DALL-E and Midjourney had been deeply mixed. Professional illustrators, photographers, and designers were alarmed — their work had been scraped to train these models without consent or compensation. Some felt their livelihoods threatened. Many were angry.

Adobe’s response was to take the longer route. Rather than training on scraped internet content, they trained Firefly exclusively on Adobe Stock images, openly licensed content, and public domain works — content where the creators had given permission for their work to be used in AI training, and in many cases would be compensated through Adobe Stock contributor royalty programmes.

Adobe Firefly launched in beta in March 2023. The message was explicit: this AI is safe for commercial use. Not “probably safe.” Not “we think we’re covered by fair use.” Safe — and indemnified. If your commercial work using Firefly generates a copyright claim, Adobe has your back legally.

Firefly isn’t just “Adobe’s DALL-E.” It’s a different product philosophy applied to a different audience with different needs. Three things genuinely distinguish it:

Firefly uses diffusion-based image generation — the same foundational approach as DALL-E and Stable Diffusion. Your text prompt is converted to a semantic representation, which guides a noise-to-image denoising process through multiple steps until a high-quality image emerges. The technical mechanics are broadly similar.

What’s different is the training data — and this shapes the model’s aesthetic more than any architectural choice.

Firefly was trained on Adobe Stock’s library — tens of millions of professional, licensed images in a very specific range of styles. Adobe Stock is a commercial photography and illustration service. Its images are, by definition, professional-grade, commercially oriented, and well-composed. They represent what professional creatives consider “good” images for commercial purposes.

This creates a distinctive output aesthetic: Firefly tends to produce images that look like premium stock photography or professional illustration. Clean, well-lit, compositionally solid, commercially polished. It’s not trying to be a fine-art tool — it’s trying to be the best tool for creating professional commercial assets.

Generative Fill in Photoshop — how it actually works. This is Firefly’s most used and most commercially significant feature. When you use Generative Fill in Photoshop, you select an area of an existing image — a patch of sky, a background, an object you want to replace — and describe what should fill it. The model generates content that matches the lighting, colour palette, perspective, and style of the surrounding image. This isn’t just image generation — it’s context-aware image generation that understands the visual environment it’s filling into. Technically, the surrounding image is encoded and passed as a conditioning signal to the diffusion process, steering it toward content that will be indistinguishable from the rest of the photo.

Generative Expand works similarly but in reverse: you extend the canvas beyond an image’s edges, and Firefly generates plausible content for the extended area — maintaining the lighting, style, and visual logic of the original. A landscape photo taken in portrait orientation can become a wide-angle panorama.

Custom Models for Enterprise. Adobe’s enterprise tier allows companies to fine-tune Firefly on their own brand assets — creating a custom model that generates images in their specific visual identity. A company can train Firefly on their product photography and brand imagery, then generate new content that is recognisably “their” aesthetic. Unlike open-source fine-tuning approaches, this is managed entirely within Adobe’s infrastructure with enterprise security and compliance guarantees.

Structure Reference and Style Reference. Firefly allows you to upload an image as a structural or stylistic reference — similar to ControlNet in Stable Diffusion, but within a managed, legally safe environment. Upload a composition reference and Firefly will generate a new image following the same spatial layout. Upload a style reference and it will apply that aesthetic to new content.

The most impactful use of Firefly for most professionals isn’t the standalone image generator — it’s Generative Fill inside Photoshop. Here’s what that looks like in real professional workflows:

Raw image quality is behind Midjourney. Firefly’s images are consistently professional but rarely stunning in the way Midjourney’s best outputs are. Training on Adobe Stock — which emphasises commercial utility over artistic ambition — produces results that look like what they are: very good stock photography. For creative concept work where aesthetic brilliance matters, Midjourney remains ahead. For commercial production where legal safety matters, Firefly wins.

The integration advantage is significant but locked to Adobe. Generative Fill in Photoshop is genuinely transformative for anyone in the Adobe ecosystem. If your workflow doesn’t involve Adobe products, Firefly’s integration advantage disappears entirely. The standalone firefly.adobe.com interface is good but not exceptional relative to competitors.

Price is a consideration. Firefly access requires an Adobe Creative Cloud subscription ($55+/month for the full suite) or generates credits that run out quickly on lower tiers. For businesses already on Creative Cloud, it’s effectively free. For those who aren’t Adobe subscribers already, subscribing just for Firefly is expensive relative to alternatives.

The training data approach is genuinely more ethical — but not perfect. Training only on licensed content is meaningfully better than scraping the internet. But questions remain: were Adobe Stock contributors truly informed about how their work would be used? Is the compensation adequate? Adobe’s approach is more transparent and fairer than most competitors, but “better than the rest” doesn’t mean “uncontroversial.”