Scale Virtual Try-On with Amazon Nova Canvas on AWS

Online fashion returns, often due to poor fit or style, cost retailers billions and increase carbon emissions. To combat this, Amazon introduces a scalable virtual try-on solution utilizing Amazon Nova Canvas on AWS. This technology targets retailers and e-commerce companies, enabling customers to visualize garments on themselves or models, significantly enhancing the online shopping experience and reducing return rates.

The core of the solution is Amazon Nova Canvas‘s virtual try-on capability, which leverages two 2D image inputs: a source image (person/living space) and a reference product image. Key features include automatic product placement via auto-masking (supporting “GARMENT” and “PROMPT” mask types for specific body parts or natural language descriptions) and manual controls for precise adjustments. The system meticulously preserves crucial details like logos, textures, and garment draping, offering comprehensive styling options. This high-fidelity, fast inference model is suitable for real-time applications and can be deployed across various channels, including e-commerce sites, mobile apps, in-store kiosks, and social media platforms.

Technically, the solution is built on a robust, event-driven AWS serverless architecture. It integrates Amazon S3 for image storage, Amazon SQS, AWS Lambda, Amazon DynamoDB for metadata and job tracking, and AWS Step Functions for workflow orchestration. Amazon Nova Canvas, accessed through Amazon Bedrock, handles mask generation and pose detection. Real-time status updates are managed via Amazon API Gateway and WebSockets, ensuring continuous user engagement with try-on requests typically completing within 7-11 seconds. This combination of generative AI and serverless services provides an accurate, scalable, and cost-effective answer to a critical retail challenge.

(Source: https://aws.amazon.com/blogs/machine-learning/building-a-scalable-virtual-try-on-solution-using-amazon-nova-on-aws-part-1/)