Personalized Poster Production and Distribution System for Video Recommendation

In the era of information overload, delivering the right content at the right time is crucial for video platforms; iQIYI’s technical product team therefore built a personalized poster production and distribution system to enhance recommendation effectiveness.

The background explains that personalized recommendation leverages user interests and behavior, and that a poster serves as the first visual impression of a video, heavily influencing user decisions. Netflix’s experiments showed that poster format, quality, and the number of people depicted significantly affect user engagement.

The system’s overall framework consists of AI‑based poster creation, offline and online distribution modules, and business strategies. AI poster production includes intelligent image cropping, smart screenshot extraction from video frames, and ZoomAI‑based quality enhancement, all powered by multiple AI algorithms for face, object, and text detection.

For distribution, an offline pipeline uploads poster assets to CDN, extracts rich visual and behavioral features, and trains ranking models (e.g., FM, DeepFM). An online service retrieves candidate posters, applies the trained models, and uses multi‑armed bandit (MAB) algorithms—both context‑free (Epsilon‑Greedy, UCB, Thompson Sampling) and context‑aware linear models—to balance exploration and exploitation.

Application strategies address poster richness (tag‑based and similarity‑based diversification), display de‑weighting for repeated exposures, a gradual retirement mechanism for underperforming posters, and scene‑matching that aligns poster tags with UI text.

Online experiments on TV and mobile showed consistent CTR improvements across placement, video, and poster dimensions, with visual examples illustrating the gains.

The article concludes by suggesting future research on generative poster synthesis, end‑to‑end ranking models that jointly consider video content and poster design, and further personalization breakthroughs.