How JD’s “JingDianDian” AI Platform Revolutionizes E‑commerce Content Creation

Background and Motivation

In e‑commerce, product image quality and marketing copy directly affect purchase decisions, but creating these assets consumes significant time and cost. JD Retail built the “JingDianDian” AIGC platform to let merchants generate high‑quality images, videos and copy with a single click.

Key Technical Challenges

Massive data processing required for training text‑to‑image models.

Precise control over image composition, style and layout for product consistency.

Ensuring accuracy and appropriate style in automatically generated marketing copy.

Rapid model optimization to keep up with fast‑changing market trends.

Core Innovations

Large‑scale Data‑driven Base Model

JD collected billions of retail images and trained a diffusion‑transformer (DiT) base model using a custom high‑throughput data pipeline, enabling the model to understand diverse product categories and scenes.

Zero‑Shot Controllable Generation

ReferenceNet injects product‑specific visual features into the generation process, while a self‑developed ControlNet provides fine‑grained control over contour, style and layout without degrading the underlying model.

Multimodal Understanding for Copywriting

A multimodal product‑understanding model extracts visual, textual and user‑review signals to build a FAB (Feature‑Advantage‑Benefit) knowledge base. Combined with a Retrieval‑Augmented Generation (RAG) pipeline, the system produces fact‑accurate, colloquial marketing copy that mitigates hallucination.

Reinforcement‑Learning‑Driven Optimization

User behavior data (clicks, conversions, feedback) continuously fine‑tunes the generation policies, and the platform dynamically allocates resources, scaling only the needed intelligent agents for high‑traffic verticals.

Performance and Business Impact

Since launch, the platform handles over 10 million AI calls per day, serving more than 350 k merchants. Reported efficiency gains exceed 95 %, reducing image‑creation time from days to seconds and cutting per‑image cost by over 99 %.

In a case study for home‑decor 2D scene generation, traditional workflows required a week and tens of thousands of yuan, whereas “JingDianDian” produced comparable results with a few uploaded photos in seconds, eliminating the need for physical shooting and post‑processing.

Resource‑efficient inference combines large‑ and small‑model joint reasoning, lowering GPU usage by up to 90 % while maintaining generation quality.

Future Directions

Further improve the base diffusion model and expand zero‑shot control capabilities.

Integrate more business signals to enable continuous self‑learning.

Extend the modular intelligent‑agent architecture to new media types such as video and audio.

Broaden AI‑plus applications across JD’s ecosystem, including fashion styling and interior design previews.