How AI Virtual Try‑On Boosted Fashion Sales by 80%: JD’s Innovative Solution

01 Introduction

In the fashion e‑commerce sector, the effectiveness of a product’s main image directly influences click‑through and conversion rates. Traditional real‑model photography is expensive, time‑consuming, and often relies on subjective experience, leading to a disconnect between content supply and sales performance.

The JD Retail Technology team launched the “JingDianDian” AI try‑on feature together with the JingMai A/B testing platform. By using AIGC to generate realistic model images and short videos, they achieved an 80%+ increase in order conversion during the 618 promotion.

02 Background

2.1 Business Scenario

Creating fashion product visuals involves complex steps such as model booking, scene setup, and post‑production, resulting in high costs and long cycles—often up to a month for a new SKU. Small merchants, constrained by budget, resort to flat or low‑quality images, which fail to attract customers.

Traditional shoots are based on studio experience and may not align with the preferences of the actual buying audience, making it difficult for fashion merchants to leverage A/B testing tools for data‑driven image optimization.

2.2 Technical Challenges

(1) Realism of AI‑generated models

Skin texture and hair details often appear overly smooth, lacking natural pores and resulting in an “oil‑slick” appearance.

Body part distortions such as incorrect limb proportions or missing fingers can occur.

(2) Garment detail and style fidelity

Difficulty accurately reproducing complex silhouettes, folds, and multi‑layered outfits.

Inadequate rendering of material textures (e.g., denim roughness, silk sheen, wool fluffiness).

Color inaccuracies leading to mismatched consumer expectations.

(3) Adaptation to different models

Challenges in fitting garments naturally on models of varying body types, causing unrealistic draping or stretching.

03 Technical Implementation

3.1 Innovations and Practices

(1) Model Library Upgrade – Built a diverse virtual model pool covering multiple ethnicities, genders, body shapes, and poses. Prompt‑engineering and a skin‑detail LoRA model enhance realism. OpenPose extracts high‑quality poses, which are stored as standard templates for garment fitting.

(2) Core Algorithms – Utilized a 12‑billion‑parameter Flux‑Fill diffusion model (Transformer‑based) for superior image generation and editing. Integrated dwpose for precise pose estimation and a parsing model for fine‑grained segmentation, ensuring pixel‑accurate garment masks.

Redux feature extraction captures garment visual attributes, which are injected as priors into the inpainting process to preserve texture, pattern, and color fidelity. A dynamic adaptive mask adjusts automatically for different garment categories, enabling consistent “one‑click” virtual try‑on.

(3) AI Video Generation – Extends static model images into 10‑15 second videos showing model rotations, enhancing visual appeal.

3.2 Results

During the 618 event, a male KA brand used JingDianDian to batch‑generate multiple style main images (e.g., business elite, casual social, mature) and corresponding videos. A/B testing compared three AI‑generated sets against a traditional manually‑produced set.

Traditional production cost per SKU: ¥2,000‑¥20,000 with a cycle ≥15 days.

AI solution cost: ¥0 with a cycle ≤0.5 day.

AI‑generated “mature stable” style outperformed the baseline, increasing order volume by over 80%.

Overall, main‑image selection efficiency improved by 30× and operational decision‑making became data‑driven.

04 Future Outlook

(1) Personalized AI Try‑On – Expand from single‑style generation to fully personalized outputs covering diverse ages, body types, skin tones, and special groups (e.g., pregnant, disabled), while simulating fabric drape and lighting.

(2) Fully Automated A/B Loop – Enable one‑click generation and experiment launch; the system automatically replaces underperforming assets with higher‑performing AI‑generated ones.

(3) Predictive AI Generation – Train conversion‑rate prediction models on historical data to proactively generate optimal fashion imagery, automatically selecting top‑performing combinations for deployment.