AWS Shows How to Build a Virtual Fitting Room for Retail on Nova Canvas

AWS published a detailed breakdown of how retailers can build an AI-powered fitting room and personalized clothing search on its cloud services. At the heart of the solution is not a new standalone product, but a ready-made reference architecture that can be deployed as a serverless application and adapted for an online store.

How the Solution Works

The architecture is built around Amazon Nova Canvas, Rekognition, Titan Multimodal Embeddings, and OpenSearch Serverless. Nova Canvas handles photorealistic fitting: a user uploads a photo, the system takes a product image and uses virtual try-on mode to generate the final image. Rekognition analyzes clothing type, body regions, and other features to ensure the mask fits correctly and the result looks believable.

All logic is split across several AWS Lambda functions: frontend, fitting processing, recommendation generation, dataset upload, and smart search. S3 and DynamoDB are used for storage, and the vector index is built in OpenSearch Serverless. AWS emphasizes that the architecture is modular: you can take just the virtual fitting room, just recommendations, or assemble the full loop.

Deployment is described almost like a ready-made recipe for a development team: clone the repository, install dependencies, build the stack through SAM, and then run guided deployment. After that, it remains to load a mini-dataset with 60+ products and separately run indexing to activate search and recommendations.

What the Store Gets

For the user, this looks like a standard shopping flow, but with an additional layer of AI. After uploading a photo, you can either try on your own item or select a product from a prepared set. In AWS's demo, it uses a dataset of 60+ fashion items, and processing the result typically takes about 15 seconds. Beyond the fitting itself, the service immediately nudges toward the next action: save the result, download it, or request similar products.

• Virtual fitting on customer photo
• Recommendations for similar products by visual similarity
• Natural language search like «blue dresses under $100»
• Filtering by color, price, category, and gender
• Analytics collection on sessions, popular products, and user actions

The recommendation layer is built on Titan Multimodal Embeddings: images and text are converted to vectors of dimension 1024, after which OpenSearch searches for nearest matches in near-real time. The search is also not limited to keywords: AWS states that the agent distinguishes at least three user scenarios—outfit selection, budget-based search, and style exploration. Additionally, DynamoDB collects analytics on fittings, unique users, popular categories, and daily activity, making the solution useful not only for the storefront but also for merchandising.

Risks and Economics

The solution has important caveats. AWS directly warns that the basic deployment does not include authentication on API Gateway, so such a template cannot be deployed to production without modifications. Separately, it recommends checking all user images: validate format, size, and resolution, and run them through moderation before sending to S3 and the generation pipeline. Otherwise, the store risks incurring not only unnecessary costs but also security and content problems.

Basic deployment does not include authentication on API

Gateway endpoints.

You need one region where Nova Canvas, Titan Multimodal Embeddings, Rekognition, and OpenSearch Serverless are simultaneously available; the article recommends us-east-1. The application itself is built through AWS SAM and deployed with a single command, and the code is published in AWS's GitHub repository. In a rough monthly estimate for workshop-level load, the most expensive component is Nova Canvas—about $60 for 1,500 fitting images.

The rest of the infrastructure, including OpenSearch, NAT Gateway, KMS, S3, and DynamoDB, adds a comparatively small amount. AWS also provides a load reference: 50 virtual fittings, 100 search queries, and 75 recommendations per day over a month of operation. For a pilot, this is sufficient to understand the economics of the scenario before a full launch.

But if the store plans to roll out such a feature to mass traffic, it will need to separately plan caching, model call limits, monitoring through CloudWatch, and abuse prevention to ensure the final unit economics don't spiral.

What This Means

AWS is betting on applied retail scenarios where generative AI impacts not the wow factor, but returns, conversion, and average order value. For stores, this is useful as a ready-made technical template: you can quickly test whether virtual fitting reduces purchase hesitation, whether natural language search drives sales growth, and whether the scenario is worth scaling further.