Adobe's Firefly Bet: A Billion Beta Images, Creative Control, and Trust at Scale

Alexandre explains that Adobe has three major business units—Creative Cloud, Document Cloud, and Experience Cloud—and AI is part of all of them. Creative Cloud includes Photoshop, Illustrator, and Adobe Express, where Adobe trains models on its homegrown Sensei platform and then tech-transfers them into products. Examples already deployed include auto-masking, content-aware fill, object detection, and neural filters in Photoshop. Adobe Research uses the Sensei platform and public clouds to let hundreds of researchers train models.

Document Cloud's Liquid Mode uses AI to reflow PDFs on mobile phones so archived papers are easier to read. In Experience Cloud, Adobe uses AI for anomaly detection in analytics, marketing content generation, and brand generation in Adobe Experience Manager. Alexandre says Adobe has been adding smaller models into workflows since 2016, then embraced large diffusion models and LLMs when quality improved. That shift led to the Firefly suite of models.

Adobe launched Firefly in March as a beta, and by the time of the conversation it had generated more than a billion images across Firefly.com, Photoshop's Generative Fill, Adobe Express, and Illustrator. Firefly was still in beta and its content was not yet cleared for commercial use, but customers were already excited about productivity and creativity. Alexandre presents the launch as the start of a broad wave of integrations rather than a single product.

When Jacob Effron asks what proportion of design AI now does, Alexandre points to Adobe's history of technological waves: digital publishing, digital photography, the internet, and mobile. In each wave, content creation became cheaper and easier and reached a larger audience, which raised concerns about job security. But demand for content increased faster than the cost of creating it fell, so Adobe's customers found opportunities to stay relevant and become more productive or creative.

Adobe believes most content will be touched by AI in one way or another, but not fully automated. Alexandre says creative people will remain in control of both creation and curation. Much of that AI will appear through editing rather than full generation: selecting an object, retouching a photo, or generating pixels inside an existing scene. The strategic question is therefore how customers adopt AI to stay relevant, not whether AI replaces them outright.

Alexandre describes Photoshop as a Boeing 747 cockpit: extremely powerful, but full of controls that require mastery. Neural Filters introduced an intent-driven editing paradigm with sliders for smile, age, or eye shape instead of adjustment layers and pixel manipulation. That interactive approach was delightful and user-friendly while still giving users control. It showed Adobe could reinvent content authoring without forcing users to learn a completely new manual process.

Prompt-based interfaces lower the entry barrier massively, even for people who do not know Photoshop, because users can express what they want in a sentence. But Adobe's customer base expects more control than a prompt alone, so the company bets on a mix of language, pointing, and sliders. Alexandre's 11-year-old daughter Rebecca tried Firefly and said the result was not what she imagined, illustrating the gap between a prompt and the image in a user's head. Adobe therefore invests in both easier human-computer dialogue and stronger control mechanisms.

Pixel-perfect Photoshop users know what beautiful design is and how to describe it, so they want mechanisms that let them create the art they intend. Alexandre says the end state of creative human-computer interaction has not been reached. When users type something, they often do not get exactly what is in their mind. Adobe's response is to combine generative capability with controls that preserve intent.

One early idea was to use an LLM to index Photoshop's help system and guide users to existing tools. The Photoshop team pushed back: a co-pilot should get users to the end result, not instruct them on a complex tool. Generative Fill operationalized that view: users select an area, AI helps with selection, and a prompt can remove a distractor, insert an object, or retouch the selection. Alexandre compares it to an Uber: users need to reach a destination without knowing how to drive.

Adobe serves four segments with different levels of creative mastery: consumers on Firefly.com, communicators through Adobe Express, creative professionals in Photoshop, Illustrator, or Premiere, and enterprises using a mix plus Workfront. Firefly.com has an inspiration stream that lets new users remix curated community creations and see the prompt, reducing cold start. Adobe's style engine offers curated lighting, composition, and style presets so users can write a simple sentence and still control the look. Changing a style can preserve the subject or emotion while changing lighting, unlike prompt changes that may lose the scene's identity.

Adobe Express templates are an entry gateway for communicators, and Adobe plans to bake generative capabilities into templates so a template can carry a prompt along with layers, effects, and background removal. For professionals, Photoshop's contextual bar appears near a selection with options such as Generative Fill; clicking Generate can use the rest of the image as an implied prompt, so users do not have to type anything. Alexandre says professionals find this dialogue about 10x easier, and Adobe is now focused on better LLM understanding of full-sentence instructions such as replace her face with blonde hair.

For enterprises, brand-guideline compliance is a major concern, especially across campaigns and sub-brands. Adobe wants generative AI to create content that is automatically compliant with brand and campaign rules, reducing QA burden and solving the content-velocity problem. In Adobe Experience Manager, Adobe is exploring contextual recommendations that autogenerate portions of text in the brand's voice. The common thread is in-context assistance rather than a blank prompt.

Alexandre says trust and safety are top of mind, starting with training data. Firefly is trained on Adobe Stock and licensed content rather than scraped internet data, which Adobe says reduces IP-infringement risk and harmful-content risk. Adobe Stock has automated and manual moderation, so illegal or harmful assets are not accepted into the marketplace. Bias remains because any marketplace dataset is biased, and training teaches the model that bias.

An internal Adobe test with tens of thousands of employees gave a healthy stress test and surfaced about 10 big problems before launch. Adobe built a person detector and job detection so prompts like lawyer do not default to Caucasian men in their 60s. It debiases outputs using country-specific skin-tone distributions plus age and gender distributions. It also uses toxicity detectors, deny and block lists, NSFW filters at the end of generation, and specific safeguards for prompts involving children.

Adobe has pushed back on claims that it trains on customer Creative Cloud data, saying it never shipped models trained on customer data. Firefly training uses Adobe Stock and additional licensed content, while customer Creative Cloud data is separated. On Firefly.com, the beta terms allow Adobe to store prompts and generated images and use them in training. Alexandre describes that as Adobe's way of doing RLHF.

That Firefly.com data supports RLHF through explicit signals such as likes and reports and implicit signals such as download, save, or share. Adobe is integrating this feedback loop into the next generation of Firefly models so reinforcement learning favors content users like and avoids content they dislike. Alexandre describes Firefly.com as Adobe's RLHF system for collecting insight into user preferences.

Firefly seemed to ship at the speed of light, but Alexandre traces it to research work that began in 2019, before diffusion models and before the Firefly name. Gavin Miller invited him to an Adobe Research retreat where a researcher presented Gen Shop, a generative Photoshop built with GANs. The quality looked like science fiction, but the vision was there, so Alexandre reached out and formed the Neural Filters team. Neural Filters became the productization arm that accelerated research models into Photoshop.

Neural Filters shipped more than 100 small models for colorization, JPEG artifact reduction, Smart Portrait, season changing, and other tasks. Many ran on-device in Photoshop using Core ML and Windows ML, which built expertise in optimized inference. When Adobe made its 2022 bet on generative AI, the team and technology stack were ready. The larger generative models run in the cloud because they are too big for on-device quality today.

Adobe Research has more than 200 researchers and welcomes 300 PhDs each summer, but research needed a productization arm. Alexandre moved under Adobe's CTO and created an applied science and machine learning group. Applied research refines models, trains the last 20% that takes 80% of time and GPU cycles, and deploys them as services. Product teams such as Photoshop, Express, and Illustrator call those APIs.

Adobe runs a horizontal layer plus vertical tech-transfer teams embedded in each product line. The Sensei team provides distributed training and curated datasets, while Gen Services creates efficient inference services; average response time is below 15 seconds for a 1K by 1K image. Product teams build user experiences on top and can submit pull requests into a managed model zoo. Alexandre calls this the AI super highway: shared infrastructure that accelerates tech transfer, with product experts deciding which workflows need AI most.

Alexandre declines to discuss pricing directly, saying he is not the right person, but explains the cost levers. The models are large and run in the cloud, so every generation carries a cost. Adobe wants best-in-class quality across images, video, vectors, and 3D, which pushes toward bigger models and encoders and higher cost. The quality bar and model size are therefore in tension with price.

Customers do not want to wait 60 seconds, so Adobe seeks an efficient frontier between quality and interactivity. Higher interactivity also tends to lower cost. Adobe is looking at distillation, pruning, quantization, open-source inference technologies, and new silicon types to keep costs in check. Commercialization details should become clearer when Firefly goes GA.

Alexandre expects software companies to become AI companies, even if they stop calling themselves AI companies because AI becomes just a new way of building software. Model architectures will keep changing, so Adobe avoids falling in love with a single model. It may ship a model, move it into maintenance, and adopt a next-generation architecture with 2x quality or performance. The pace of innovation in the space is exponential.

Ten years out, Alexandre expects AI to replace more of the software stack and enable massive changes in what content is created. Enterprises face a content-velocity and content-supply-chain problem: they want individualized communication but cannot afford humans to create every variant. Generative AI could produce infinite content personalized to a segment of one, with opt-in data about how a viewer prefers information. That would let brands and small businesses communicate differently with constituents.

Creative Cloud could evolve from a tool for visual designers into a platform for creative directors of hyper-personalized meta experiences. Human-computer interaction should become more fluid, less convoluted, and accessible to larger audiences. Alexandre hopes the AI becomes so core that it is not even mentioned. The Creative Suite would enable content creation with higher creativity and velocity.