Adobe has spent nearly two decades integrating AI across its product suite, and with the launch of Firefly, it is now deploying large-scale generative models to over a billion image generations in beta alone. Alexandru Costin, VP of Generative AI and Sensei at Adobe, explains how Adobe is weaving AI into every layer of its business, from Photoshop to enterprise content management, while navigating copyright, bias, and the future of creative work.
Adobe’s AI journey and product landscape
Adobe operates three major business units: Creative Cloud (Photoshop, Illustrator, Adobe Express), Document Cloud (Acrobat), and Experience Cloud (content management, analytics, CRM).
AI is embedded across all three:
In Creative Cloud, AI powers features like auto-masking, content-aware fill, neural filters, and now generative fill in Photoshop.
In Document Cloud, AI enables “liquid mode” in Acrobat, which reflows documents for easier reading on mobile devices.
In Experience Cloud, AI handles anomaly detection in analytics and brand-compliant content generation in Adobe Experience Manager.
Adobe’s internal AI infrastructure is called the Sensei platform, a homegrown distributed training environment running on public clouds that supports hundreds of researchers.
In 2022, Adobe made a strategic bet on large models, both diffusion models and LLMs, leading to the creation of the Firefly suite of models, launched in March 2023.
Firefly has already generated over a billion images during its beta phase across firefly.com, Photoshop (generative fill), Adobe Express, and soon Illustrator.
How AI is changing creative workflows
Adobe’s strategic view is that AI will touch most content creation but will not fully automate it; creative people will remain in control of both creation and curation.
Historically, each wave of technological disruption (digital publishing, digital photography, internet, mobile) made content cheaper and easier to create, which raised concerns about job security, but demand for content grew even faster, creating more opportunity for creators.
Adobe sees the same pattern with AI: content creation becomes more accessible, demand grows exponentially, and creators who embrace AI become significantly more productive.
Even before generative AI, most Photoshop users were already using AI through selection models and GANs; generative fill extends this by allowing users to generate or edit pixels through natural language prompts.
Designing AI into complex products
Photoshop has accumulated decades of features and controls, making it powerful but complex to master.
Adobe’s early AI work included neural filters, which introduced intent-driven editing: instead of manually manipulating pixels, users could adjust sliders for things like smile intensity, age, or head direction, with GANs handling the pixel-level changes.
When prompt-based interfaces emerged, Adobe recognized they dramatically lower the entry barrier but also realized that professional users want more control than a prompt alone can provide.
The design philosophy Adobe is pursuing combines language, pointing, and sliders, acknowledging that the ideal human-computer interaction for creativity has not yet been reached.
The most successful workflows Adobe has observed are “co-pilot” experiences: users select an object (often AI-assisted), describe what they want in a prompt, and the system does the work. This mirrors the Uber analogy: you don’t need to drive, you just need to say where you want to go.
Generative fill in Photoshop exemplifies this: users select an area, type a prompt (or leave it blank), and the system generates content that matches the surrounding image.
Onboarding and educating users across segments
Adobe serves four customer segments, each with different levels of creative mastery:
Consumers: Use firefly.com for quick results without learning complex tools. Adobe provides an inspiration stream of curated community creations that users can click on, see the prompt, and riff from. A style engine lets users pick styles, lighting, and composition from curated options rather than writing complex prompts. Prompt auto-completion models are also in development.
Communicators/Marketers: Use Adobe Express, which relies on templates. Adobe plans to bake generative capabilities directly into templates so users can start creating immediately.
Creative Professionals: Use Photoshop, Illustrator, Premiere, etc. For them, Adobe introduced a contextual bar that appears near any selection, offering generative fill without requiring the user to navigate the UI. Leaving the prompt blank uses the rest of the image as an implied prompt. Professionals tend to write full sentences (e.g., “replace her face with blonde hair”), so Adobe is investing in better LLM understanding of these natural language instructions.
Enterprises: Use a mix of tools plus workflow management (e.g., Workfront). Adobe plans to introduce contextual AI recommendations, such as auto-generating text in Experience Manager that respects brand voice.
A key enterprise concern is brand compliance: companies need generated content to automatically respect brand and campaign guidelines. Adobe is innovating to solve the “content velocity” problem, enabling rapid content creation that is brand-compliant out of the gate.
Trust, safety, and bias mitigation
Adobe trains Firefly exclusively on Adobe Stock content, not on scraped internet data, to reduce IP infringement risk and give artists consent, control, and compensation.
Adobe Stock has a rigorous content moderation process (both automated and manual), which means the training data is already filtered for harmful and illegal content.
Despite this, training data from any marketplace carries inherent bias. Adobe discovered bias issues during an internal stress test with tens of thousands of employees in February 2023, just before the March launch.
To address this, Adobe built a person detector that identifies when a prompt references a person (often through a job title like “lawyer”) and applies debiasing to ensure fair representation across skin tones, genders, and age groups, calibrated to the country of origin of the request.
For harm reduction, Adobe uses toxicity detectors, deny lists, block lists, and NSFW filters. Special protections were added for prompts involving children, such as preventing associations with tobacco.
Adobe maintains a feedback mechanism on firefly.com and in Photoshop so users can report issues, providing new training data points to continuously refine the models.
How customer feedback improves the models
Adobe does not train on customer data stored in Creative Cloud; Firefly is trained on Adobe Stock plus some museum and public domain content.
On firefly.com, Adobe’s terms of service allow it to store prompts and generated images and use them for training, which serves as Adobe’s reinforcement learning from human feedback (RLF) system.
Explicit signals (likes, reports) and implicit signals (downloads, saves, shares) are logged and used to teach future model generations what users prefer.
Adobe is building a feedback loop into the next generation of Firefly models so that reinforcement learning steers generation toward content users like and away from content they dislike.
Building and deploying at scale
Firefly’s development began years before its 2023 launch, starting with GANs in Adobe Research as early as 2019 under the name “generative Photoshop.”
A dedicated neural filters team was created to accelerate tech transfer from research to product, shipping over 100 small AI models and building expertise in distributed training and optimized inference.
When Adobe made its strategic bet on generative AI in 2022, the team, technology stack, and infrastructure were already in place.
Adobe runs inference in the cloud because current models are too large for on-device use at the required quality level, though some neural filters do run on-device using Core ML and similar frameworks.
Adobe has built custom orchestration and computing layers for efficient imaging pre- and post-processing, achieving average response times below 15 seconds for a 1K×1K image.
The rapid January-to-March 2023 timeline was possible because the foundational work had been ongoing for years, with strong executive support from the CEO and president of digital media.
How Adobe structures its AI teams
Adobe uses a hybrid model combining horizontal and vertical teams:
Horizontal layer: Adobe Research (200+ researchers, 300 summer PhDs) operates under the CTO and focuses on modality-specific labs (image, video, vector, design, 3D). The Sensei team provides distributed training infrastructure and curated datasets. An applied research team refines models for production, and a Gen Services team deploys them as efficient APIs.
Vertical layer: Each major product line (Photoshop, Express, Illustrator) has a tech transfer subgroup of applied researchers who take the APIs and build user experiences around them, because they understand the core workflows and customer segments best.
Adobe calls this architecture an “AI Super Highway,” a managed model zoo where product teams can submit pull requests and build workflows on top of shared services, accelerating innovation and tech transfer.
Pricing and cost considerations
Running large generative models in the cloud carries significant costs, driven by model size, parameter count, and encoder complexity needed for high-quality output.
Adobe is investing in distillation, pruning, quantization, and exploring new silicon types to keep costs manageable.
The company is balancing the tradeoff between quality and interactivity: customers won’t wait 60 seconds for a generation, so finding the efficient frontier between speed, cost, and output quality is a core challenge.
Pricing details for Firefly post-beta have not been finalized, but cost management is a key focus for the team.
Adobe’s 10-year vision for Creative Cloud
Adobe believes software companies will become AI companies, with AI replacing increasing portions of the traditional software stack.
Model architectures will continue to evolve rapidly; Adobe avoids locking into any single architecture and iterates to each new generation that offers meaningful quality or performance gains.
The biggest change Adobe anticipates is hyper personalization: enterprises currently face a content supply chain bottleneck where it is too expensive to create individualized content for every customer segment. Generative AI will enable infinite, personalized content at scale, potentially down to the “segment of one,” where a website serves images and text tailored to each individual viewer (opt-in).
Creative Cloud will evolve to enable customers to transition from visual designers to creative directors of meta-experiences that are hyper-personalizable, changing how brands and small businesses communicate with their audiences.
Human-computer interaction will become more fluid, less convoluted, and accessible to much larger audiences, with AI becoming so embedded that it may no longer be explicitly mentioned as a feature.
