Adobe Firefly hit the scene in March 2023, a family of generative AI models tailored for creative professionals. Firefly's creators took a deliberate stance on training data, opting for licensed content over scraped internet data, a move aimed squarely at addressing copyright concerns surrounding Stable Diffusion and Midjourney.
The training data Adobe chose for Firefly is telling. They drew from Adobe Stock images, public domain content, and openly licensed material. The goal was clear: a model whose outputs wouldn't raise eyebrows over derivative works from copyrighted images in the training set. For enterprise creative teams and agencies that need to produce commercially licensed content, this distinction is significant.
Adobe's curation of licensed content for Firefly involved combing through Adobe Stock's 150 million+ images, public domain repositories, and Creative Commons-licensed assets. The team used automated tools like Tika and Apache NiFi to extract metadata and validate licenses, but manual audits were still required for edge cases. This process took months and required partnerships with legal teams to avoid overreaching. The result was a dataset of ~200 million images, but the trade-off was a narrower stylistic range compared to models trained on uncurated web data. Firefly often fails to replicate niche art styles or hyper-specific cultural references not well-represented in Adobe's curated sources.
Firefly isn't a standalone AI tool; it's AI capability woven into the very fabric of Creative Cloud applications like Photoshop, Illustrator, and Adobe Express. Generative fill in Photoshop lets you select a region and describe what to put there, replacing or extending image content with AI-generated pixels that match the surrounding style and lighting. This integration is key to its appeal.
Adobe's engineering team optimized generative fill for real-time performance using TensorRT and ONNX Runtime, achieving sub-second latency on NVIDIA GPUs. They also implemented a two-stage rendering pipeline: a low-res preview generated instantly, followed by a high-res render in the background. This approach kept Photoshop responsive during interactive sessions, but it required careful memory management to avoid GPU thrashing on consumer-grade machines. Lavkesh recalls a production incident where a team's custom build failed to handle 4K canvas sizes, causing Photoshop to crash repeatedly during large-scale batch edits.
Adobe's model for Firefly credits is a departure from the unlimited use offered by competitors. Enterprise plans include a volume of generative credits, which creates a different user behavior. Users treat generations as a resource to spend deliberately rather than iterating infinitely. Whether this approach yields better outputs or simply introduces friction is up for debate.
The credit model mirrors how Adobe charges for cloud rendering or stock image downloads, but it's a double-edged sword. While it ensures predictability for finance teams, it forces creative teams to budget for AI usage. Lavkesh has seen this pattern before in tools like AWS SageMaker, where metered billing leads to "AI hourglass" workflows—teams stockpile credits before deadlines and exhaust them in bursts. Adobe mitigates this with tiered pricing and bulk credit discounts, but workflow disruptions still occur when teams hit their monthly caps during critical sprints.
The professional creative community's reaction to Adobe Firefly is more nuanced than their response to Stable Diffusion or Midjourney. The training data choice reduces the ethical concern, and the integration into existing tools minimizes workflow disruption. Many professional designers report using generative fill for specific tasks (background extension, object removal and replacement) while producing the primary creative work through traditional methods. Firefly augments the workflow rather than replacing it.