Bark.com et AWS s'associent pour construire une solution de génération vidéo à grande échelle

This post is cowritten with Hammad Mian and Joonas Kukkonen from Bark.com. When scaling video content creation, many companies face the challenge of maintaining quality while reducing production time. This post demonstrates how Bark.com and AWS collaborated to solve this problem, showing you a replicable approach for AI-powered content generation. Bark.com used Amazon SageMaker and Amazon Bedrock to transform their marketing content pipeline from weeks to hours. Bark connects thousands of people each week with professional services, from landscaping to domiciliary care, across multiple categories. When Bark’s marketing team identified an opportunity to expand into mid-funnel social media advertising, they faced a scaling problem: effective social campaigns require high volumes of personalized creative content for rapid A/B testing, but their manual production workflow took weeks per campaign and couldn’t support multiple customer segment variations. If you’re facing similar content scaling challenges, this architecture pattern can be a useful starting point. Working with the AWS Generative AI Innovation Center, Bark developed an AI-powered content generation solution that demonstrated a substantial reduction in production time in experimental trials while improving content quality scores. The collaboration targeted four objectives: Production time – Reduce from weeks to hours Personalization scale – Support multiple customer micro-segments per campaign Brand consistency – Maintain voice and visual identity across generated content Quality standards – Match professionally produced advertisements In this post, we walk you through the technical architecture we built, the key design decisions that contributed to success, and the measurable results achieved, giving you a blueprint for implementing similar solutions. Solution overview Bark collaborated with the AWS Generative AI Innovation Center to develop a solution that could tackle these content scaling challenges. The team designed a system using AWS services and tailored AI models. The following diagram illustrates the solution architecture. The solution architecture consists of the following integrated layers: Data and storage layer – Amazon Simple Storage Service (Amazon S3) stores assets including training data, generated video segments, reference images, and final outputs. Model artifacts and custom inference containers are stored in Amazon Elastic Container Registry (Amazon ECR). Processing layer – AWS Lambda orchestrates the multi-stage pipeline, with AWS Step Functions managing the workflow state across the seven-step generation process. Amazon Bedrock with Anthropic’s Claude Sonnet 3.7 handles text generation tasks, including customer segmentation, story generation, and quality evaluation. GPU compute layer – To serve Wan 2.1 Text2Video-14B reliably, we run a multi-GPU inference container that shards the model across eight GPUs on a single p4de.24xlarge SageMaker instance using tensor parallelism. TorchServe fronts the endpoint for request handling, and torchrun launches one worker process per GPU. We use Fully Sharded Data Parallel (FSDP) sharding—a technique for splitting the model components across GPUs—for the text encoder and the diffusion transformer to stay within GPU memory limits without offloading weights to CPU. Because video diffusion is long-running, the endpoint is tuned with an extended inference timeout and a longer container startup health-check window to accommodate model load time and help avoid premature restarts. Amazon Elastic Container Service (Amazon ECS) containers on GPU-enabled g5.2xlarge instances handle speech synthesis for narrator voice generation, scaling to zero during idle periods. User interface layer – A React frontend with Amazon Cognito authentication provides a video studio interface where marketing teams can review, edit, and approve generated content through natural language commands. Creative ideation pipeline Now that you understand the overall architecture, let’s examine how you can implement the creative ideation pipeline in your own environment. The pipeline transforms customer questionnaire data into production-ready storyboards through three stages. Stage 1: Custom er segment generation The pipeline begins by analyzing Bark’s customer questionnaire data using Amazon Bedrock with Anthropic’s Claude Sonnet 3.7. The large language model (LLM) processes survey responses to identify distinct customer personas with structured attributes including demographics, motivations, pain points, and decision-making factors. For example, in the domiciliary care category, the system identified segments such as: The Overwhelmed Family Caregiver – Adults in their 40s–50s balancing work responsibilities with caring for aging parents, prioritizing reliability and trust The Independence-Focused Senior – Elderly individuals seeking to maintain autonomy while acknowledging the need for occasional assistance Each segment