Yan Li

I currently lead the Multimodal AI and AIGC initiatives at JD.com, overseeing product strategy, operations, design, algorithm development, and engineering R&D. On the algorithm side, my team focuses on advancing e-commerce multimodal understanding models, controllable visual generation models, and unified models tailored for online retail scenarios. Through our product efforts, we launched JD Oxygen Vision, an AIGC-powered platform designed to help merchants create, test, and optimize product visuals and advertising materials. The platform now serves a wide range of sellers on JD.com. We’ve also introduced a virtual try-on capability that enables customers to visualize clothing fit and receive personalized outfit recommendations. Looking forward, we’re working toward the next-generation content paradigm in e-commerce: fully personalized, AI-generated visual materials for each user. By delivering unique product visuals tailored to individual preferences, we aim to enhance the shopping experience while driving better conversion and business growth for our merchants.

As Tech Lead at Kuaishou’s Kling AI Team, I led the development of data infrastructure for collecting, cleaning, and managing multi-billion-scale high-quality image/video datasets. My team engineered the foundational text-to-image generative model (Kolors), innovating core components including multilingual text encoders, variational autoencoders (VAE), and diffusion model architectures. Leveraging this proprietary technology, we developed ID-preserving controllable image generation and e-commerce AIGC solutions—such as virtual try-on capabilities. To date, our Kolors model has generated over 340 million premium images for diverse applications, serving millions of users while ensuring consistent brand identity in generated assets.

At Kwai Global, I led multimodal content understanding across all scenarios — including short videos, live streams, e-commerce products, and commercial ad creatives — for regions spanning Latin America (Portuguese/Spanish), Southeast Asia (Indonesian/Urdu), and the Middle East (Arabic/Turkish). This required developing multilingual models to handle linguistic and cultural nuances in 10+ languages. Additionally, I applied AIGC capabilities to repurpose existing videos through automated editing and localized dubbing, enriching Kwai’s overseas content ecosystem. I also generated commercial ad creatives (e.g., synthetic product demos, regionally adapted visuals) to support user acquisition campaigns. My work established a scalable framework for multimodal content understanding and generation across diverse global markets.​

At WeChat Video Channels, I led the development of core content understanding capabilities, including video classification from broad categories to fine-grained tags, video duplicate detection algorithms, and facial recognition systems. These multimodal models formed the foundational features for the recommendation system, providing critical signals for new content during cold-start phases.​

As a core AI researcher at Tencent WeChat, I led the development of early-stage facial recognition algorithms deployed across WeChat’s ecosystem, enhancing security and user authentication for hundreds of millions of users. My work included optimizing algorithms for accuracy in challenging scenarios like cross-age verification, contributing to industry-leading performance on benchmarks. For WeChat Moments, I spearheaded multimodal content understanding systems, designing a comprehensive knowledge framework with 40,000+ semantic tags to analyze image-text relationships and improve content recommendations. Additionally, I drove AI innovation for WeChat 7.0’s flagship feature—AI-powered video soundtracking—developing algorithms that dynamically matched music to video content, which became one of the industry’s most acclaimed video enhancement tools during its release. These projects leveraged computer vision, multimodal learning, and generative AI to elevate user experiences at scale.