Abstract: In this talk, we explore how to choose the most suitable image embedding models for real-world e-commerce applications such as visual search, product clustering, or duplicate detection. Based on a large-scale benchmark conducted at Adevinta (Leboncoin), we compare 20 state-of-the-art models (CLIP, DINOv2, ViT, etc.) across multiple domains and tasks. Full results are available in our paper: https://arxiv.org/abs/2504.07567, and we’ll share key insights to help practitioners make informed, task-specific model choices.