What Would You Click? Personalized Video Thumbnail Generation with Preference-aware Highlight Retrieval

Traditional video platforms often display a single, generic thumbnail for all users, failing to account for the diverse interests of individual viewers. This paper addresses the challenge of personalized video thumbnail generation, asking how to create thumbnails that are both tailored to specific user preferences and visually coherent with the original video content.

The researchers propose a two-stage framework that bridges user preference modeling with controllable image generation:

1. Personalized Highlight Retrieval: This stage uses a cross-attention mechanism to model fine-grained interactions between user history and video frames. A hierarchical gated fusion module then balances these user-specific preferences with global video semantics (derived from titles and tags) to identify the most relevant visual anchors.

2. Personalized Thumbnail Generation: The framework uses a Vision-Language Model (VLM) to convert the retrieved anchors into structured editing instructions. These instructions guide a diffusion-based model to perform personalized edits on a base frame. The process includes an editing intensity predictor to calibrate the degree of change based on user history and a verify-and-retry loop to ensure the final output remains faithful to the original video.

Experiments on two public datasets demonstrate that the proposed framework outperforms existing retrieval-based and generative baselines in terms of thumbnail quality, content fidelity, and style personalization. A user study confirms that these personalized thumbnails significantly improve click-through preferences, validating the real-world utility of the approach for enhancing user engagement.

By shifting from a one-size-fits-all approach to a generative, user-centric paradigm, this work provides a scalable way to optimize the visual entry points of video content. It demonstrates that combining deep user-preference modeling with modern generative AI can effectively increase user engagement on video platforms.