How AI‑Generated Virtual Try‑On Boosted Fashion Sales by 80%+

1 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 creating a gap between product presentation and consumer perception.

JD Retail Technology’s "JingDianDian" AI (model) try‑on, combined with the JingMai A/B testing platform, uses AIGC to generate realistic model images and short videos, automatically validated by A/B experiments, achieving an 80%+ increase in order conversion during the 618 promotion.

2 Background

2.1 Business Scenario

Fashion product material creation involves model booking, scene setup, shooting, and post‑production, leading to high costs and long cycles (up to a month). Small merchants often resort to flat or low‑quality model 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 merchants from using A/B testing to validate different main‑image versions.

"JingDianDian" ( https://ai.jd.com/ ) offers AI try‑on that can batch‑generate diverse real‑model images, ensuring seamless fit of clothing folds, lighting, and body shape, eliminating the "photoshopped" feel. AI video generation creates 10‑15 s dynamic model videos, turning static images into motion videos. Integrated with JingMai A/B testing, brands can quickly optimize conversion during large promotions.

2.2 Technical Challenges

(1) Realism of AI‑generated models

"Oiliness": AI images often have overly smooth skin lacking pores and blemishes, appearing artificial.

Body part distortion or missing parts (e.g., incorrect finger count, facial distortion).

(2) Clothing detail and style fidelity

Pattern and silhouette: AI struggles with complex styles such as voluminous skirts, layered outfits, or accurate length proportions.

Texture: Difficulty reproducing realistic denim roughness, silk sheen, wool fluffiness.

Color: Color shifts can mislead consumers.

Fine details: Buttons, zippers, lace patterns, logos may be inaccurate or blurred.

(3) Adaptation to different models

Natural fitting: Clothing may not drape naturally on varied poses.

Body type compatibility: AI may stretch tight clothing on larger bodies or make loose clothing look oversized on slim bodies.

3 Technical Practice

3.1 Innovations and Implementation

To address the challenges, "JingDianDian" AI try‑on introduced several innovations.

(1) Model library upgrade

A high‑quality virtual model library covering multiple ethnicities, genders, body types, and poses was built. Large‑scale prompt engineering and LoRA fine‑tuning improve skin realism. OpenPose extracts high‑quality poses, which are stored as standard templates for clothing try‑on.

The system intelligently matches models to clothing style, scene, and target audience, also considering color harmony between garment and skin.

(2) Core algorithms

The system centers on a 12‑billion‑parameter Flux‑Fill diffusion model (Transformer‑based). It integrates dwpose for pose estimation and a parsing model for fine‑grained segmentation, achieving sub‑pixel keypoint accuracy and precise region masks.

A Redux feature extractor captures multi‑dimensional visual features from input garments; after normalization these features guide the inpainting process, preserving texture, pattern, and color.

A dynamic adaptive mask adjusts automatically to garment categories (top, bottom, dress) during training, ensuring stable, realistic try‑on across styles.

Examples of AI try‑on results:

3.2 Practice Results

Case study: a male‑apparel KA brand during the 618 promotion.

The brand used AI to batch‑generate multiple main‑image styles (business, casual, mature) and 10‑15 s model videos. A/B testing compared three AI‑generated sets plus one manually created set.

First round elimination: traditional studio images lagged 66.67% in sales within 72 h.

Final victory: AI‑generated "mature" style outperformed, boosting sales by over 80%.

Operational decision evolution: data‑driven selection improved main‑image selection efficiency by 30×.

Traffic agility: 24/7 automatic optimization maximized conversion during the promotion.

Key metrics

Cost per SKU main‑image/video : Traditional ¥2,000‑20,000 and ≥15 days; AI ¥0 and ≤0.5 days (100% cost reduction, >95% efficiency gain).

Main‑image A/B testing : Traditional ≥30 days, AI ≤1 day with batch generation (30× faster).

Technical integration complexity : Traditional required coordination of design, modeling, shooting; AI can be operated by a junior staff member online (95% simplification).

3.3 Reasons for Breakthrough

(1) Cross‑team collaborative innovation

Joint effort of "JingDianDian", "JingMai", and fashion category operations built an integrated AIGC + A/B testing pipeline.

(2) Algorithmic architecture innovation

Adoption of the large‑scale Flux‑Fill model solved detail loss and pose distortion.

(3) Data‑closed‑loop iteration

A/B experiment feedback (click‑through, conversion) feeds back to improve model generation (e.g., optimal age‑group and lighting combinations).

(4) Engineering optimization

Task scheduling, smart prompting, and precise garment classification reduce latency while maintaining quality.

(5) Asset reuse

Historical real‑model photos are converted into a digital asset library, enabling reuse and cutting material costs by over 80%.

4 Future Outlook

(1) Personalized AI try‑on for millions of users

Extend from single‑style generation to full‑dimensional personalization (age, body shape, skin tone, special groups) and simulate fabric drape and lighting.

(2) Fully automated A/B experiments

Move from manual upload to one‑click automatic experiment launch, with the system auto‑replacing under‑performing assets based on data.

(3) Predictive AI try‑on generation

Train conversion‑rate prediction models on historical data to proactively generate optimal main‑image combinations, e.g., predicting top‑selling outfit pairings for specific demographics.