How Alibaba’s DiDa Platform Uses AI to Automate Fashion Outfit Matching

Preface

Since the "LuBan" AI designer produced 8,000 posters per second, the team explored richer visual content generation, naming the effort "DiDa" (滴搭). DiDa is a platform that uses deep learning to create fashion outfit combinations and accompanying text, supporting millions of pairings for tens of thousands of creators.

1. DiDa Platform Overview

DiDa unifies several subsystems—front‑end selection, algorithmic engine, image composition, and personalized delivery—into a single platform. It ingests product images, titles, and operational knowledge, retrieves compatible items from a million‑scale catalog, generates descriptive titles, and formats the result for visual display. The platform has been deployed across multiple Alibaba businesses since 2017, serving millions of items and tens of thousands of merchants.

2. DiDa Algorithms

2.1 Image Matching Algorithm

The image pipeline uses a CNN (Inception‑v3) to extract high‑dimensional product vectors, followed by K‑means clustering with category constraints to obtain refined embeddings. Two matching strategies are explored:

Sequence‑based generation using an LSTM to model the ordering of items.

Unordered generation using a Deep Aggregated Network (DAN) that treats item sets as combinations.

Both approaches share a DSSM (Deep Semantic Similarity Model) component that aligns product vectors based on positive samples from high‑CTR logs and expert‑curated outfits, and negative samples from low‑CTR logs.

2.2 Text Generation Algorithm

Text generation is framed as a summarization task using a Context‑aware Pointer‑Generator Network (CPGN). The encoder processes product titles, tags, and side information; the decoder generates copy‑or‑generate words, balancing between copying from the source and generating new tokens via a soft switch. Coverage mechanisms prevent repeated words.

3. DiDa Engineering Platform

The platform runs on XTensorflow (XTF), a distributed TensorFlow training and online scoring system built on Porsche Blink. It enables daily model updates for CNN, LSTM, DAN, and CPGN at a massive scale, providing real‑time inference for image scoring, context‑graph queries, and DSSM retrieval.

Additionally, an intelligent layout engine combines designer templates with learned composition rules to produce aesthetically pleasing multi‑product collages.

4. Business Cases

4.1 iFashion

DiDa supplies outfit images and descriptions for the iFashion scene, mixing algorithmic and creator‑generated content to improve cost efficiency and conversion.

4.2 Mobile Taobao Homepage Focus

By generating multi‑product focus images, DiDa enriches the visual experience and boosts CTR and UCTR by double‑digit percentages.

4.3 Good Goods Channel

DiDa refines long creator titles into concise, information‑dense copy, enhancing user decision‑making.

5. Future Work

Develop an end‑to‑end model that integrates image processing directly into the generation pipeline.

Expand personalization capabilities and explore more user‑generated content scenarios.