Florence

Florence is Microsoft's computer vision foundation model, published in November 2021. It expands visual representations from coarse (scene) to fine (object), from static images to dynamic video, and from RGB to multiple modalities such as captions and depth, trained on web-scale image-text data.

The model achieved state-of-the-art results across 44 representative benchmarks, including 83.74% top-1 zero-shot accuracy on ImageNet-1K and strong transfer on COCO detection, VQA, and Kinetics action recognition. It adapts with minimal customization through fine-tuning, linear probing, or few-shot transfer.

Florence later shipped as the vision backbone of Azure AI Vision services, powering production captioning, retrieval, and video summarization. The foundation weights themselves were not open-sourced.