Case Study — Generative AI

Virtual Try Room —
AI Garment Try-On for Retail

An in-store virtual try-on system powered by a 4-model AI pipeline: OpenPose and DensePose map the body, MediaPipe extracts precise measurements, and Stable Diffusion XL Inpaint renders photorealistic garment overlays in under 3 seconds.

Generative AIComputer Vision

Build a Similar System Back to Projects

AI Pipeline

📷

INPUT

User Photo

→

🦴

STEP 1

OpenPose + DensePose

→

📏

STEP 2

MediaPipe Measurements

→

✨

STEP 3

Stable Diffusion XL

→

👔

OUTPUT

Virtual Try-On

<3sEnd-to-End Processing

4AI Models in Pipeline

SDXLStable Diffusion Version

100%GDPR Compliant

Overview

Reinventing the In-Store Fitting Room with Generative AI

Physical fitting rooms are a major source of friction in retail: long queues, hygiene concerns post-pandemic, and chronic size mismatches leading to high return rates. Existing virtual try-on tools are built for e-commerce — they lack the speed and accuracy needed for in-store use. Virtual Try Room was designed specifically for in-store kiosks and retail environments.

The Problem

Four Retail Pain Points

Fitting room queues — peak-hour congestion reduces throughput and customer satisfaction
Hygiene concerns — post-pandemic reluctance to try on shared garments
Size uncertainty — inconsistent sizing across brands drives high return rates
Stock limitations — customers leave when their size isn't available in the fitting room

The Solution

4-Model AI Pipeline

A purpose-built AI system for in-store use: customers capture a photo at a kiosk or via mobile. OpenPose + DensePose map body posture and anatomical landmarks. MediaPipe extracts precise measurements (chest, sleeve, shoulder, waist). Stable Diffusion XL Inpaint renders a photorealistic garment overlay. A custom ML model recommends the right size from store inventory.

AI Pipeline

Four Models, One Seamless Experience

The pipeline is designed for real-time operation: each model is specialized for its role, and the full workflow completes in under 3 seconds on GPU-enabled edge hardware.

Pose & BodyOpenPose detects 18 body keypoints; DensePose creates dense UV body surface maps for precise anatomical landmark identificationOpenPoseDensePoseDetectron2

MeasurementMediaPipe Pose extracts 33 landmarks; custom geometry pipeline converts pixel distances to real-world measurements in cm/in with body depth estimationMediaPipeNumPyOpenCV

Size EngineCustom ML model maps extracted measurements to brand-specific sizing charts; cross-references with real-time inventory to suggest available sizesPyTorchScikit-learnStore API

Garment RenderStable Diffusion XL Inpaint with prepare_mask_and_masked_image pipeline; VAE + CLIP encoders; UNet denoising for photorealistic garment drapingSDXLDiffusersCUDAONNX

Post-ProcessOpenCV image alignment and edge blending; lighting normalization to match garment to ambient conditions; final compositing and preview generationOpenCVPILFastAPI

Key Features

Designed for Retail Speed and Accuracy

Body Pose Mapping — 18-keypoint OpenPose skeleton detection + DensePose UV surface maps for accurate anatomical feature localization.

Anthropometric Measurement — MediaPipe extracts chest circumference, sleeve length, shoulder width, and waist — converted from pixels to cm/in.

Photorealistic Overlay — Stable Diffusion XL Inpaint generates garment overlays that respect fabric draping, body contours, and lighting conditions.

Body Segmentation — Detectron2-powered body-object segmentation separates the user from background for clean garment placement.

Size Recommendation — Custom ML model trained on brand-specific sizing charts recommends the optimal size with confidence score.

Real-Time Inventory Sync — Checks in-store inventory availability before recommending a size; suggests alternatives if the item is out of stock.

GDPR Compliant — No biometric data retained after session. All processing is in-memory; output previews are not linked to identity.

Edge-Optimized — Models exported to ONNX and optimized for NVIDIA edge devices; designed to run on in-store kiosk hardware.

Technical Highlights

Engineering Challenges Solved

Measurement Pipeline

Pixel-to-Cm Conversion

MediaPipe's 33 3D landmarks are used with a reference shoulder-to-shoulder baseline to establish a pixel-per-cm ratio. Sleeve length is computed as the geodesic path from shoulder landmark 11/12 → elbow → wrist. Chest is computed as twice the horizontal distance from the spine to the armpit landmark.

Garment Rendering

StableDiffusionXLInpaintPipeline

The custom prepare_mask_and_masked_image() function creates a precise garment-region mask from DensePose UV coordinates, then passes it along with the latent noise and text prompt to the SDXL inpainting UNet. VAE decodes the final latent to a photorealistic RGB output.

Technology Stack

State-of-the-Art Computer Vision Stack

PythonFastAPIPyTorchStable Diffusion XLDiffusersOpenPoseDensePoseMediaPipeDetectron2OpenCVONNXCUDANumPyReactGradioCOCO DatasetDeepFashion2

Built By

Development Team

Fatima Abbas

Developer

Mohsin Sabir

Developer

Amber Razzaq

Developer