Intelligent Creative Content Ecosystem for JD Advertising: Content Understanding, Generation, and Distribution

The presentation introduces the rapid growth of content ecosystems as a key driver of internet development and AI adoption, focusing on JD's advertising business and its intelligent creative platform.

Background and Problem Definition : The rise of short‑video and content‑driven e‑commerce creates a need for better content understanding and recommendation to support advertising.

Intelligent Creative Definition : AI‑based content understanding, generation, and distribution form a three‑stage pipeline that enables personalized ad creatives.

Content Understanding System : Includes three parts—basic algorithmic tagging, material admission (filtering and compliance), and quality/aesthetic evaluation. Techniques such as scene classification, semantic segmentation, product detection, OCR, and aesthetic scoring are applied to billions of SKUs.

Material Admission : A multi‑modal intelligent review system combines data annotation, automated cleaning, and similarity‑based retrieval to filter inappropriate or low‑quality assets.

Quality and Aesthetic Evaluation : A model‑based end‑to‑end solution learns weights across multiple aesthetic dimensions and incorporates click‑through‑rate signals.

Intelligent Content Generation : Covers video summarization (shot detection, tagging, relevance scoring), image generation (scene‑aware GANs, component‑wise rendering, real‑time engine), and copy generation (VAE‑enhanced Transformer models with weak supervision). All generated assets are directly used in ad placements.

Creative Selection & Distribution : A multimodal ranking model aligns user, product, and creative features, using an explore‑exploit (EE) strategy to balance novelty and relevance. The model incorporates user behavior, creative embeddings, and CTR prediction.

Business Impact : The system powers billions of daily generated creatives across core ad slots, improves click‑through rates by over 30% in some cases, and has earned JD's technology innovation award.

Q&A Highlights : Discussed aesthetic scoring methodology, copy‑generation reliability, component‑based image generation, label taxonomy creation, and offline evaluation metrics (AUC, revenue lift).