What is Florence-2?

It is a general-purpose VLM (Visual Language Model) that can handle multiple tasks such as caption generation, object detection, segmentation, and OCR with a single model by looking at an image.

In this page, we will focus on four tasks often used in ComfyUI: "Caption Generation", "Object Detection (Coordinate Extraction)", "OCR", and "Q&A about Images".

Custom Node

Florence2Run Node

Florence2Run is the main node for having Florence-2 execute tasks on the input image. By switching task, you can use functions such as caption generation, object detection, and OCR properly.

caption, detailed caption

Generates a natural language caption from the image.

florence2_detailed_caption.json
  • caption
    • Explains the outline of the image simply.
  • detailed caption
    • Explains the composition and appearance in a little more detail.

However, if the purpose is only "caption for prompts", using a caption-specific model such as JoyCaption will produce much more flexible and high-quality results.

caption_to_phrase_grounding

Outputs the position of the object in the form of a rectangle (bounding box) for each phrase of the specified caption.

florence2_caption_to_phrase_grounding.json
  • It is characterized by being able to take the position even with slightly complex instructions such as "left tree" or "red car".
  • By extracting coordinates with the 🟨 Florence2 Coordinates node and combining it with a segmentation model such as SAM2, you can use it to mask only specific objects.

ocr

Reads characters in the image and outputs them as text.

florence2_ocr.json

docvqa

A task to answer questions about the image.

florence2_docvqa.json
  • You can ask questions like "Where is XX in this image?" or "What is the value of this table?" and receive the answer in text.
  • Imagine usage similar to throwing an image to ChatGPT and asking questions.