How AI-Powered Virtual Try-On Boosted Fashion Sales by 80%+

1 Introduction

In fashion e‑commerce, main‑image quality directly impacts click‑through and conversion rates. Traditional photo shoots with real models are costly and slow, and rely on subjective experience, causing a gap between product and effect.

JD Retail’s technology team launched “JingDianDian” AI virtual try‑on combined with the JingMai A/B testing platform. Using AIGC, it generates realistic model images and short videos for apparel, achieving an 80%+ increase in order conversion during the 618 promotion.

2 Background

2.1 Business Scenario

Fashion apparel requires complex material production involving model booking, set construction, shooting, and post‑processing, leading to high costs and long cycles (up to a month per SKU). Small merchants often resort to flat or low‑quality images, reducing main‑image appeal.

Traditional studio shoots produce images that may not match the preferences of the target audience, and the high cost prevents many merchants from using A/B testing to validate image performance.

“JingDianDian” ( https://ai.jd.com/ ) offers AI‑driven try‑on that can batch‑generate diverse, real‑person‑level model images and short dynamic videos, eliminating the “photoshopped” feel and enabling rapid SKU‑level optimization.

2.2 Technical Challenges

The AI try‑on faces several challenges:

Realism of generated models : skin texture, hair detail, and lighting often appear overly smooth (“oil‑slick” effect) and lack natural imperfections.

Body part errors : incorrect limb proportions, missing fingers, facial distortion.

Garment detail and style fidelity : difficulty reproducing complex silhouettes, textures (denim, silk, wool), and accurate colors.

Fit to different models : unnatural draping, incorrect folds, and poor adaptation to varied body shapes.

3 Technical Practice

3.1 Innovations

To solve the above, “JingDianDian” introduced:

Iterative virtual model library : multi‑ethnicity, gender, body shape, and pose models built with large‑scale prompt engineering and LoRA fine‑tuning. OpenPose extracts high‑quality poses for reuse.

Core algorithm : a 12‑billion‑parameter Flux‑Fill diffusion model (Transformer‑based) combined with dwpose pose estimation and parsing for sub‑pixel keypoint accuracy. Redux extracts garment features as priors for in‑painting, ensuring texture and pattern consistency.

Dynamic adaptive mask : automatically adjusts mask generation per garment category, maintaining stable, realistic try‑on across styles.

Examples of generated results are shown below.

3.2 Results

In a case study with a male‑apparel KA brand during the 618 event, AI‑generated main images and 10‑15 s videos for multiple styles (business, casual, mature) were produced automatically. A/B testing compared three AI‑generated sets plus a traditional manually‑produced set.

Key metrics comparison:

Cost per SKU image/video : Traditional ¥2,000‑¥20,000, ≥15 days; AI ¥0, ≤0.5 day → cost ↓100 %, efficiency ↑95 %.

Main‑image A/B test turnaround : Traditional ≥30 days, manual; AI ≤1 day, batch → efficiency ↑30×.

Technical integration complexity : Traditional requires coordination of design, model, shooting; AI can be operated with a single click by an intern → barrier ↓95 %.

Key outcomes:

First‑round elimination: traditional images lagged AI by 66.67 % in conversion within 72 hours.

Final victory: AI‑generated “mature” style boosted conversion by over 80 %.

Operational decision making shifted from experience‑based to data‑driven, improving image selection efficiency by 30×.

24/7 automated optimization maximized traffic conversion during promotions.

3.3 Breakthrough Reasons

Cross‑team collaborative innovation between “JingDianDian”, “JingMai”, and fashion operations.

Algorithmic architecture using the large‑scale Flux‑Fill diffusion model to eliminate detail loss and pose distortion.

Data‑closed‑loop: A/B test results feed back into the generative model for continuous improvement.

Engineering optimizations balance compute cost and generation quality.

Resource reuse: historic real‑model photos are transformed into a digital asset library, cutting reuse cost by >80 %.

4 Future Outlook

Personalized AI try‑on for thousands of user profiles (age, body type, skin tone, special groups).

Fully automated A/B experiments with intelligent decision loops.

Predictive AI generation that proactively creates top‑performing main‑image combos based on conversion forecasts.