Overview of Baidu's Wànxiàng System for Large‑Scale Rich Media Processing

With the rapid growth of rich‑media content (images, videos, audio) on the Internet, traditional web‑page‑centric search faces new challenges. Users now consume information through native apps and short‑video platforms, resulting in diverse content formats and fragmented feedback signals.

The Baidu search engine therefore needs to handle massive amounts of multimedia data, extract semantic features, and aggregate user behavior across multiple devices and applications.

Wànxiàng (meaning “all‑encompassing”) is Baidu’s internal system designed to process, index, and retrieve rich‑media assets at scale. It supports all image and video processing required by Baidu’s products, managing billions of media entities and delivering billions of processing requests daily.

The system is built around three core subsystems:

Qianren (Blades) : extracts low‑level and high‑level features from individual media items (e.g., OCR, scene detection, clarity, tags). It orchestrates tens of thousands of CPU cores and GPU/FPGA resources, converting all feature calculations into DAG jobs for efficient scheduling and deduplication.

Chuyu (Initial) : analyzes relationships between media entities (e.g., duplicate, similar, containment, event‑based clustering). It performs fingerprint‑level comparisons to discover cross‑media links such as video clips derived from the same source.

Danding (Athanors) : stores and aggregates feature data, merging attributes of identical or similar entities so that downstream retrieval can operate on content‑level signals rather than page‑level signals.

Wànxiàng also includes auxiliary services for cropping, transcoding, and editing. Its design emphasizes two key metrics:

Scalability : the ability to process petabytes of images and videos, leveraging heterogeneous resources (CPU, GPU, FPGA) across hundreds of thousands of cores.

Timeliness : delivering processed features, indexes, and quality signals within product iteration cycles to ensure up‑to‑date search relevance. By converting rich‑media into structured semantic attributes and aggregating user feedback at the content level, Wànxiàng enables Baidu to support image search, video search, recommendation, and other rich‑media‑driven services with high accuracy and low latency.