Gan.AI Enables Real-Time Creation of Personalized Videos to Drive Brand Engagement

Gan.AI launched in 2021 with a bold mission: to democratize personal video production with generative artificial intelligence (generative AI). The founders wanted to empower everyone, from singular real-estate agents to global enterprises, to easily create hyper-personalized messages for targeted, meaningful marketing. Today, even sports teams such as the San Antonio Spurs and Mumbai Indians use Studio by Gan.AI—the company’s main product—to drive fan engagement.

In its initial iteration, Studio was utilized by large consumer brands. Companies would provide a list of recipients to target, and Gan.AI would use this data to generate a custom video for each one, including personalized audio, lip-sync, and visuals. This process could take up to 4 minutes per video.

Then, a new self-serve use case emerged for real-time videos at scale. Anupreet Singh, chief revenue officer at Gan.AI, explains, “Some of our largest customers realized if their users could create their own videos, they could attract people who were otherwise not as engaged with the brand. It was a good opportunity for them to build the top of the marketing funnel and collect data to further engage potential customers.”

Gan.AI faced the challenge of significantly reducing latency in video creation. “Every second matters for consumers visiting a brand’s site to make a video; even 2 minutes was too long,” Singh says. Inferencing for its generative AI models was the primary bottleneck in Gan.AI’s infrastructure. The company sought an efficient, low-latency solution to scale its deep learning models.

Gan.AI had already been using Amazon Web Services (AWS) for several years, including using Amazon Elastic Container Service (Amazon ECS) with AWS Lambda to host its machine learning models on CPUs. Following discussions with its AWS account team, the company chose Amazon SageMaker to inference its deep learning models on GPU instances. In particular, the use of multi-model endpoints would offer the speed required for large-scale projects in real time.

Suvrat Bhooshan, chief executive officer at Gan.AI, says, “Amazon SageMaker helped us host multi-model endpoints without a high cold-start latency of 5−6 minutes. Now, we can serve multiple content requests within 30−40 seconds and scale without much engineering effort.” 

Gan.AI spent nearly three months in testing, and with each challenge it encountered along the way, AWS was there to help. For example, when the team needed to ensure quick auto scaling of multi-model endpoints, AWS provided guidance on architecting in such a way that a single GPU instance could be used to serve up to five requests concurrently, responding within a few seconds. Gan.AI typically deploys 8−10 instances that are always running, and that number can scale to about 200 instances when required.

“New users of Amazon SageMaker should thoroughly test and experiment with its various features and different GPU instance types before implementation,” Bhooshan advises. “For example, we found that offloading non-GPU processes to AWS Lambda enabled more efficient utilization of Amazon SageMaker resources. This was especially useful for scaling in a business such as ours with a mix of deep learning and CPU-intensive workloads.”