Real-time thumbnail selection improves short video engagement globally
Billion-Scale Thumbnail Optimization for Uncurated Short-Form Videos via Multi-Armed Bandits
Machine Learning
Summary
Short videos often don’t have customized thumbnail images chosen by their creators, which can make it hard for viewers to decide what to watch. The authors built a system that automatically picks better thumbnail images in real time for billions of short videos worldwide. They use a smart method that tries different images and quickly learns which ones get more viewer attention, guided by a model that predicts image quality. This approach improved important user engagement measures when used on a major video platform.
What this means in practice
- •For video streaming platforms: Automatically select appealing thumbnails for short videos to increase viewer discovery and engagement at large scale.$Commercial implications: Enables platforms to monetize more videos by improving user interaction with dynamically optimized thumbnails.
- •For content recommendation engineers: Incorporate dynamic thumbnail optimization into recommendation pipelines to enhance content click rates without manual curation.
Authors
Ying Han, Ling Liu, Fabio Soldo, Vu Nguyen, Danio Wang, Liz Kidd, Yongle Cao, Theodore Rose, Su-Lin Wu, Romer Rosales
Abstract
This paper introduces a real-time thumbnail optimization system deployed at a global $O(B)$ scale on a major short-form video platform. Unlike traditional long-form content, where custom thumbnails are heavily curated by creators, a considerable fraction of short-form videos are published without human-selected artwork. To address this uncurated corpus, we present a fully automated, end-to-end framework that replaces static default frames with dynamic, data-driven selections across billions of videos. To the best of our knowledge, this is the first published work demonstrating an online Multi-Armed Bandit framework successfully deployed at an $O(B)$ scale for uncurated short-form video discovery. Our solution pairs a multi-stage candidate generation pipeline with a low-latency serving infrastructure. By initializing the exploration framework with image-specific priors derived from a deep visual quality model, the system minimizes exploration costs and dynamically serves optimal thumbnails at serving time. Global deployment demonstrates statistically significant improvements in core user discovery and engagement metrics.