Runs Florence-2 (Microsoft, ~0.23B) entirely in your browser via Transformers.js. Unlike the embedding lab (which turns an image into one vector), a VLM produces text: it can caption an image, read the text in it (OCR), and locate objects. Drop a screenshot of a webpage and hit Read text to watch it read.
Drop an image here, or click to choose
Florence-2 is a sequence-to-sequence model: a DaViT vision encoder turns the image into a grid of tokens, and a small language decoder reads those tokens plus a task prompt (like <OCR>) to write the answer. That token-reading step is why it can transcribe a page a fixed-resolution embedding model cannot. First run downloads ~0.35 GB (cached afterwards). Model: Florence-2-base-ft.