How AI Virtual Try‑On Boosted Fashion Sales by 80%: A Technical Deep‑Dive

1. Introduction

In fashion e‑commerce, the main image directly influences click‑through and conversion rates. Traditional photo shoots with real models are costly and time‑consuming, and rely on subjective experience, often causing a gap between product appeal and sales performance. JD Retail’s technology team launched the "JingDianDian" AI virtual fitting solution combined with the JingMai A/B testing platform, using AIGC to generate diverse, realistic model images and short videos, achieving an 80%+ increase in order conversion during the 618 promotion.

2. Background

2.1 Business Scenario

Fashion product material creation involves model recruitment, scene setup, shooting, and post‑production, leading to high costs and long cycles (up to a month per SKU). Small merchants often resort to flat or low‑quality model images, resulting in weak main‑image attraction. Traditional studio shoots also suffer from a mismatch between studio aesthetics and actual consumer preferences, making A/B testing of main images difficult.

2.2 Technical Challenges

The AI fitting system faces several challenges:

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

Body part defects : Incorrect limb proportions, missing fingers, or facial distortions can occur.

Garment detail and style fidelity : Complex styles (e.g., layered skirts) and textures (denim, silk, wool) are hard to reproduce accurately.

Color accuracy : Color shifts reduce consumer trust.

Fine‑grained garment details : Buttons, zippers, lace, logos, and patterns may be misplaced or blurred.

Fit across different body types : The system struggles to adapt garments to diverse body shapes, leading to unrealistic stretching or loose fits.

3. Technical Practice

3.1 Innovations and Practices

To address the above issues, "JingDianDian" implemented several innovations:

Model library upgrade : Built a high‑quality virtual model pool covering various ethnicities, genders, body types, and poses. Used large‑scale prompt engineering and a LoRA model for realistic skin, combined with OpenPose for pose extraction.

Core algorithm : Adopted a 12‑billion‑parameter Flux‑Fill diffusion model (Transformer‑based) with dwpose pose estimation and parsing for sub‑pixel keypoint accuracy. Integrated a Redux feature extractor to inject garment visual features into the inpainting process, preserving texture, pattern, and color.

Dynamic adaptive mask : Adjusts mask generation based on garment category (top, bottom, dress) to maintain stable, realistic fitting across styles.

3.2 Results

In a case study with a male‑apparel KA brand during the 618 event, AI‑generated main images and 10‑15 second model videos were produced automatically. A/B testing compared three AI‑generated image/video sets plus one manually created set. Findings:

Traditional studio images lagged behind AI variants, delivering 66.67% lower transaction volume within 72 hours.

The "mature" AI style outperformed the baseline, increasing sales by over 80%.

Data‑driven selection improved main‑image optimization efficiency by 30×.

24/7 automated tuning maximized traffic conversion.

Cost and efficiency gains:

Metric

Traditional Solution

JingDianDian AIGC Solution

Improvement

SKU image/video production cost

¥2,000‑¥20,000, >15 days

¥0, ≤0.5 day

Cost ↓100%, efficiency ↑95%+

Main‑image A/B testing

≥30 days, manual design

≤1 day, batch generation

Design speed ↑30×

Technical integration complexity

Multiple teams (design, photography, location)

One‑click platform operation

Barrier ↓95%+

3.3 Reasons for Breakthrough

Cross‑team collaboration : Joint effort between "JingDianDian", "JingMai", and fashion operations created an end‑to‑end AIGC + A/B testing pipeline.

Algorithmic innovation : Leveraged state‑of‑the‑art diffusion model to solve detail loss and pose distortion.

Data‑closed loop : A/B test metrics feed back into model fine‑tuning, e.g., optimal age‑group and lighting combinations.

Engineering optimization : Task scheduling, intelligent prompting, and garment‑type recognition balance latency and quality.

Asset reuse : Historical real‑model photos are converted into a digital asset library, reducing new material costs by >80%.

4. Future Outlook

Personalized AI fitting for thousands of user profiles (age, body shape, skin tone, special groups).

Fully automated A/B experiments with AI‑driven decision loops.

Predictive AI generation: train conversion‑rate prediction models to proactively create top‑performing main‑image combos.