Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents: 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →
ReleaseInsight · On RadarAWS Machine Learning Blog→ original

AWS shows how to build a virtual fitting room for retail on Nova Canvas

AWS has shown how to assemble a serverless solution for fashion retail with virtual try-on, recommendations and smart search. The design uses Nova Canvas, Rekognition, Titan Multimodal Embeddings and OpenSearch Serverless. In the demo, a try-on takes about 15 seconds, and the package includes a dataset of 60+ products and basic analytics. But AWS explicitly warns that the template must not go into production without authentication and validation of user-uploaded images.

AI-processed from AWS Machine Learning Blog; edited by Hamidun News
AWS shows how to build a virtual fitting room for retail on Nova Canvas
Source: AWS Machine Learning Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceAWS Machine Learning Blog ↗
Editor-in-Chief
Key metric #1Virtual try-on on the customer's photo
Key metric #2Recommendations of similar products based on visual similarity
Key metric #3Natural-language search such as "blue dresses under $100"

AWS has published a detailed walkthrough of how a retailer can build an AI fitting room and personalized clothing search on its cloud services. At the center 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 design is built around Amazon Nova Canvas, Rekognition, Titan Multimodal Embeddings and OpenSearch Serverless. Nova Canvas handles the photorealistic try-on: the user uploads a photo, and the system takes an image of the garment and generates the final picture using its virtual try-on mode. Rekognition analyzes the garment type, body areas and other features so the mask is placed correctly and the result looks plausible.

The logic is split across several AWS Lambda functions: frontend, try-on processing, recommendation generation, dataset loading and smart search. S3 and DynamoDB handle storage, and the vector index is built in OpenSearch Serverless. AWS stresses that the architecture is modular: you can take only virtual try-on, only recommendations, or assemble the full pipeline.

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

What the store gets

For the user, this looks like an ordinary shopping flow with an extra AI layer. After uploading a photo, they can either try on their own item or pick a product from a prepared set. In the demo, AWS uses a dataset of 60+ fashion items, and processing usually takes about 15 seconds. Beyond the try-on itself, the service immediately nudges the user toward a next action: save the result, download it or request similar products.

  • Virtual try-on on the customer's photo
  • Recommendations of similar products based on visual similarity
  • Natural-language search such as "blue dresses under $100"
  • Filtering by color, price, category and gender
  • Analytics on sessions, popular products and user actions

The recommendation layer is built on Titan Multimodal Embeddings: images and text are converted into vectors of dimension 1024, after which OpenSearch finds the closest matches in near real time. Search is not limited to keywords either: AWS writes that the agent distinguishes at least three user scenarios — outfit selection, budget-based search and exploration of new styles. In addition, DynamoDB collects analytics on try-ons, unique users, popular categories and daily activity, which makes the solution useful not only for the storefront but also for merchandising.

Risks and economics

The solution comes with important caveats. AWS explicitly warns that the basic deployment does not include authentication on API Gateway, so the template must not be pushed to production without further work. It also recommends checking all user images: validating format, size and resolution, and running them through moderation before sending them to S3 and the generation pipeline. Otherwise the store risks not only extra costs but also security and content problems.

The basic deployment does not include authentication on API Gateway endpoints.

To run it, you need a single region where Nova Canvas, Titan Multimodal Embeddings, Rekognition and OpenSearch Serverless are all available at once; the article recommends us-east-1. The application is built with AWS SAM and deployed with a single command, and the code is published in an AWS GitHub repository.

In the sample monthly estimate for a workshop workload, the most expensive component is Nova Canvas — about $60 for 1,500 try-on images. The rest of the infrastructure, including OpenSearch, NAT Gateway, KMS, S3 and DynamoDB, adds a comparatively small additional sum. AWS also gives a load benchmark: 50 virtual try-ons, 100 search queries and 75 recommendations per day over a month. For a pilot, that is enough to understand the economics of the scenario before a full launch.

But if a store plans to take this mechanism to mass traffic, it will have to plan separately for caching, model call limits, monitoring through CloudWatch and abuse protection, so that unit economics do not spiral.

What this means

AWS is betting on practical retail scenarios in which generative AI affects not the wow factor but returns, conversion and average order value. For stores, this is useful as a ready-made technical template: they can quickly test whether virtual try-on reduces hesitation before purchase, whether natural-language search lifts sales, and whether it is worth scaling the scenario further.

Story

Key Questions Answered

How the solution works?

The design is built around Amazon Nova Canvas, Rekognition, Titan Multimodal Embeddings and OpenSearch Serverless. Nova Canvas handles the photorealistic try-on: the user uploads a photo, and the system takes an image of the garment and generates the final picture using its virtual try-on mode. Rekognition analyzes the garment type, body areas and other features so the mask is placed correctly and the result looks plausible.

What the store gets?

For the user, this looks like an ordinary shopping flow with an extra AI layer. After uploading a photo, they can either try on their own item or pick a product from a prepared set. In the demo, AWS uses a dataset of 60+ fashion items, and processing usually takes about 15 seconds. Beyond the try-on itself, the service immediately nudges the user toward a next action: save the result, download it or request similar products.

What are the key details on Risks and economics?

The solution comes with important caveats. AWS explicitly warns that the basic deployment does not include authentication on API Gateway, so the template must not be pushed to production without further work. It also recommends checking all user images: validating format, size and resolution, and running them through moderation before sending them to S3 and the generation pipeline. Otherwise the store risks not only extra costs but also security and content problems.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…