Florence-2

Instead of loading a separate network for every vision job, Florence-2 uses one sequence-to-sequence model and picks the task from a short token at the start of the prompt. Microsoft published a 0.23 billion parameter base and a 0.77 billion large version, each with a fine-tuned twin, trained on FLD-5B, roughly 5.4 billion machine-made labels spanning 126 million photos. Output arrives as text that a bundled helper converts into boxes, polygons, labels or prose, covering detection, region captions, phrase grounding, referring segmentation, OCR with optional region boxes, and captions at three levels of detail. The original checkpoints load in Transformers with trust_remote_code, and converted copies under florence-community use the native Florence2 classes now in the library. Being this small, they run on a laptop GPU or even a CPU, which is why ComfyUI nodes and dataset captioning scripts lean on them. Code and weights are MIT licensed, so commercial use is fine. Open-ended visual questions are better left to large multimodal LLMs; this model shines at structured, repeatable labeling.