AI Virtual Try‑On Transforms Fashion E‑Commerce, Raising Conversion 80%

Introduction

In fashion e‑commerce, the main product image directly influences click‑through and conversion rates. Traditional photo shoots with real models are costly, time‑consuming, and rely on subjective experience, often causing a mismatch between content supply and sales effectiveness.

Background

JD Retail’s technology team launched the “JingDianDian” AI virtual try‑on solution in partnership with the JingMai A/B testing platform. By using AIGC technology to generate realistic model images and short videos for fashion items, the solution achieved an 80%+ increase in order conversion during the 618 promotion.

Technical Challenges

Realism of AI‑generated models : Skin texture, hair details, and lighting often appear overly smooth (“oil‑slick” effect), and body parts may be distorted or missing.

Garment detail and style fidelity : Complex silhouettes, layered outfits, and fabric textures (denim, silk, wool) are difficult to reproduce accurately.

Fit across different models : Adapting garments to diverse body shapes without unnatural stretching or gaps remains challenging.

Technical Practice

The system introduced several innovations:

Model library upgrade : Built a diverse virtual model pool covering multiple ethnicities, genders, body types, and poses. Prompt engineering and LoRA fine‑tuning improved skin realism. OpenPose extracted high‑quality pose templates for consistent garment fitting.

Core algorithm : Adopted a 12B‑parameter Flux‑Fill diffusion model (Transformer‑based) combined with dwpose pose estimation and parsing segmentation. Redux feature extraction ensured garment texture, pattern, and color consistency during inpainting.

Dynamic mask adaptation : During training, masks automatically adjust to garment categories (tops, bottoms, dresses) to maintain stable, realistic try‑on results.

Results

For a major men’s fashion brand during the 618 event, AI‑generated main images and 10‑15 second videos were produced in batch for multiple style scenarios (business, casual, mature, etc.). A/B testing compared three AI‑generated variants with one traditional manually‑produced variant.

Conversion uplift : The “mature” style AI images achieved an 80%+ increase in order conversion compared with the baseline.

Cost reduction : Production cost dropped from ¥2,000‑¥20,000 per SKU with a 15‑day cycle to near‑zero cost within half a day.

Efficiency gain : Main‑image design time reduced from ≥30 days to ≤1 day, and overall workflow efficiency improved by more than 30×.

Future Outlook

The roadmap includes:

Personalized AI try‑on : Generate garments tailored to individual age, body shape, skin tone, and special groups (e.g., pregnant women, people with disabilities).

Fully automated A/B experiments : One‑click generation and automatic replacement of the best‑performing assets based on real‑time metrics.

Predictive AI generation : Use conversion‑rate prediction models trained on historical data to proactively create the most effective fashion visuals.