Adobe's Journey to Generative AI Leadership Using AWS

Rather than rely on open source models, Adobe decided to train its own foundational generative AI models tailored for creative use cases. "We want to be in control of the quality and capabilities our models need to have," explained Costin. A key differentiator was training Firefly generative AI models on licensed content, such as Adobe Stock images, and public domain content where copyright has expired.

In just nine months, Adobe launched  the Firefly family of generative AI models. These models were rapidly integrated into Adobe's flagship creative software products like Photoshop, Illustrator, Substance 3D, and Adobe Express. "The Generative Fill capability in Photoshop powered by Firefly is now the most used feature," highlighted Costin. By building on AWS, Adobe could accelerate AI training rather than building from scratch, enabling the company to focus on its differentiators and not the heavy lifting.

Adobe created an AI superhighway on AWS to build an AI training platform and data pipelines to rapidly iterate models. Adobe built its solution with Amazon Elastic Compute Cloud (Amazon EC2) P5 and P4d instances powered by NVIDIA GPUs, Amazon Elastic Kubernetes Service (Amazon EKS), Amazon Elastic Block Store (Amazon EBS), and Amazon Elastic Fabric Adapter (EFA). Adobe also used Amazon Simple Storage Service (Amazon S3) to serve as the data lake and primary repository for the vast troves of data. Adobe used Amazon FSx for Lustre high-performance file storage, for fast access to data and to make sure GPU resources are never left idle. Adobe continuously monitors and synchronizes data across multiple sources, ensuring that its models are trained on the most up-to-date and relevant information. With Firefly, Adobe does continuous testing to mitigate against any harmful biases and provides user feedback mechanisms to remediate concerns in real-time, fostering transparency and accountability.

While AI and data play a pivotal role in Adobe's creative endeavors, the company recognizes that true innovation often stems from the human touch. To bridge the gap between machine learning and human creativity, Adobe has embraced a groundbreaking technique known as Reinforcement Learning with Human Feedback (RLHF). Through RLHF, Adobe collects invaluable feedback from its vast user base, capturing likes, dislikes, downloads, and other interactions from its various applications. This wealth of human-generated data is then fed back into the training process, helping Adobe's AI models to learn and adapt, continuously refining its ability to generate assets that resonate with customers' preferences and desires. The result is a cycle of continuous improvement, where human creativity and machine learning work in tandem, each informing and enhancing the other. As Adobe's models become more attuned to the nuances of human taste and aesthetics, the assets the models generate become increasingly compelling, captivating, and tailored to the needs of the diverse user base.

As demand spiked, Adobe scaled up training by 20x in six months, aided by Amazon EC2 Reserved Instances. They implemented distributed training techniques like pipeline parallelism to efficiently utilize tensor cores. In addition to training, Adobe deployed the Firefly models for inference on AWS, implementing stateless systems across regions. They optimized inference performance and cost, exploring options like AWS Inferentia2 chips. Adobe demonstrated the ability to rapidly scale and shift hardware for optimal price and performance.