Cross-modal Image Recommendation for News Articles by Multimodal Foundation Models-based Retrieval-Reranking

Retrieving relevant images for a given news article is challenging and can be considered a special version of the cross-modal retrieval problem. This notebook paper presents our solution for the MediaEval NewsImages 2025 benchmarking task. We propose a retrieval-reranking solution based on multimodal foundation models such as VLMs and multimodal LLMs, and utilizing multiple levels of textual granularity. We report the official results of our submitted runs and additional experiments we conducted internally to evaluate our runs.