TensorFlow Lite Applications and UI2Code at Xianyu (Idle Fish)

In previous articles, Xianyu (Alibaba's second‑hand platform) revealed its work on UI2Code and Flutter at Google I/O 2018. This article shares Xianyu’s experience with TensorFlow, especially TensorFlow Lite, presented at the Google Developer Conference.

TensorFlow has become the most widely used machine‑learning framework since its 2015 release. At the conference, new features such as eager execution, standardized interchange formats, API cleanup, and post‑training quantization for TensorFlow Lite were highlighted.

Xianyu built a custom TensorFlow Lite framework for its client, focusing on extensibility, security, and dynamism. Custom OPs enable independent business‑specific models, encrypted model headers protect data, and server‑side model/OP updates allow online upgrades and gray‑scale testing.

Application 1 – Video Cover: Replaces rule‑based cover selection with an AI model that can be updated dynamically, reducing maintenance cost.

Application 2 – Video Fingerprint: Uses TensorFlow Lite as a computation framework, chaining custom OPs to form a flexible end‑to‑end pipeline that can be updated per OP or as a whole.

Application 3 – Smart Rental: Deploys a popular SSD model to recognize furniture in rental photos, automatically tagging listings and enriching search/recommendation data.

The “UI2Code” project converts a screenshot directly into UI code. Unlike Microsoft’s Sketch2Code, which aims at design prototyping, UI2Code targets pixel‑perfect, production‑ready code that developers can fine‑tune.

TensorFlow is used in UI2Code for image understanding (detecting text, objects, etc.) to drive the conversion pipeline.

The tool is currently internal to Xianyu, with plans to open‑source later pending compliance reviews.

Overall, Xianyu’s participation at the conference showcased practical AI deployments on mobile clients and invited developers to join their innovative team.