How Kuaishou Elevates Short‑Video Quality and AI Performance at NVIDIA GTC 2023

Video Quality Evaluation and Enhancement

Kuaishou processes tens of millions of UGC short videos daily, and to deliver clearer visuals each video passes through multiple stages of capture, editing, and server‑side processing, all constrained by network, bandwidth, and device factors.

Sun Ming, head of audio‑video image algorithms, introduced the KVQ video quality assessment system and the KRP/KEP enhancement pipeline, which together significantly improve perceived clarity on the consumption side.

The KVQ framework, driven by AI, was built on extensive internal test sets and iteratively refined to handle content, processing, and codec diversity, achieving commercial deployment in StreamLake and services for several well‑known companies.

Digital Human Live Streaming and Interactive Solutions

Jian Weihua presented Kuaishou’s 3D digital‑human live‑streaming platform built on the KMIP virtual world interaction framework and the KVS virtual broadcasting assistant.

In gaming scenarios, digital‑human avatars act as live hosts, guiding users through gameplay while allowing real‑time interaction, resulting in over 50% revenue growth for participating streamers and a two‑fold increase in live‑room payment rates.

Custom Performance‑Optimization Framework

Men Chunlei described an end‑to‑end sub‑graph optimization framework based on NVIDIA TensorRT, which automatically analyzes ONNX graphs, trims performance‑bottleneck sub‑graphs, and generates optimized TensorRT plugins via an AI compiler.

Performance Optimization for Large Recommendation Models

Liang Xiao explained how Kuaishou balances CPU and GPU workloads on a single server to accelerate massive recommendation models, moving critical workloads to GPU, deeply optimizing CPU algorithms, and caching data on GPU to reduce DRAM accesses.

This approach raised GPU utilization from around 20% to nearly 90% and increased throughput by more than tenfold.

Accelerating Multimodal Large Models for Short‑Video Scenarios

Zhang Shengzhuo, Han Qingchang, and Li Jie presented a comprehensive solution for multimodal massive models, covering hybrid parallel training, inference optimization, and deployment, enabling low‑cost, high‑impact applications across recommendation, advertising, search, and e‑commerce.

The work addresses three major challenges—long training time, low inference efficiency, and complex deployment—by improving model compute efficiency and ease of deployment.