Supported tasks

TaskTypes enumerates every Florence-2 task this wrapper recognises. Each one produces a FlorenceResults; which fields are populated depends on the task. The table below is the cheat-sheet.

Quick reference

Tasks that need a textInput

Most tasks ignore the third argument to Run. These three require it:

• CAPTION_TO_PHRASE_GROUNDING — textInput is the natural-language phrase to ground. Example: "the red car".
• OPEN_VOCABULARY_DETECTION — textInput is the open-vocabulary class to detect. Example: "traffic light".
• REGION_TO_* tasks — textInput encodes the region using <loc_xxx> tokens in Florence-2's 1000-bin coordinate space. See OCR for how to compute them from pixel coordinates.

For everything else, pass null or just omit the argument.

The result types

public class FlorenceResults
{
    public LabeledOCRBox[]        OCRBBox;       // OCR_WITH_REGION
    public string                 PureText;      // captions, plain OCR, region descriptions
    public LabeledBoundingBoxes[] BoundingBoxes; // OD, grounding, region proposal
    public LabeledPolygon[]       Polygons;      // segmentation tasks
}

public class LabeledBoundingBoxes
{
    public BoundingBox<float>[] BBoxes;
    public string               Label;
}

public class LabeledOCRBox
{
    public Coordinates<float>[] QuadBox;  // four corner points
    public string               Text;
}

public class LabeledPolygon
{
    public string                   Label;
    public List<Coordinates<float>> Polygon;
    public List<BoundingBox<float>> BBoxes;
}

Coordinates are in pixel space of the original image. The wrapper de-normalises Florence-2's <loc_…> tokens for you.

Picking the right task

• "What's in this picture?" → start with CAPTION, escalate to DETAILED_CAPTION or MORE_DETAILED_CAPTION.
• "Read this receipt / sign / screenshot." → OCR for plain text, OCR_WITH_REGION if you also want bounding boxes.
• "Where are the cars / people / cats?" → OPEN_VOCABULARY_DETECTION with a class prompt, or OD for generic detection.
• "Highlight the part of the image that matches this sentence." → CAPTION_TO_PHRASE_GROUNDING.
• "Describe the object inside this box." → REGION_TO_DESCRIPTION with a <loc_…> textInput.

For full code examples task by task, see the Guides.

A note on REFERRING_EXPRESSION_SEGMENTATION

REFERRING_EXPRESSION_SEGMENTATION is enumerated for completeness but does not produce valid output in the current release — the generated <loc_…> token stream is not well-formed. Use CAPTION_TO_PHRASE_GROUNDING to get a bounding box, then call REGION_TO_SEGMENTATION for a polygon as a workaround.