Object Detection

Object Detection Nodes

Object detection nodes let you identify, locate, and segment objects in images. Use loader nodes to configure each model’s weights, then connect them to the corresponding detector or segmentation node.

---

Load Florence-2 Model Loader

Inputs

None

Widgets

• Model Name: Select a Florence-2 checkpoint to load.

Outputs

• Florence-2 Model: Model object passed to Florence-2 Object Detection.

Description

Downloads and loads Microsoft Florence-2 model weights. The loaded model is passed directly to the Florence-2 Object Detection node. Florence-2 is a universal vision model capable of detection, captioning, OCR, and more — all driven by a text prompt.

Managing Files

• Directorymodels

  • Directoryflorence2

    • florence-2-base-ft
    • florence-2-large-ft

---

Florence-2 Object Detection Detector

Inputs

• Image: Input image.
• Model: Florence-2 Model from Load Florence-2 Model node.

Widgets

• Text Prompt: Comma-separated object classes or caption request.
• Task: The vision task to perform.

Outputs

• BBoxes: Bounding box coordinates.
• Labels: Detected class labels or captions.
• Image(s): Annotated output image.

Description

A universal vision node powered by Microsoft Florence-2. Unlike YOLOX which is limited to 80 COCO classes, Florence-2 is open-vocabulary — you describe what you want to detect in plain text and it finds it. It also supports captioning, OCR, and dense region proposals.

Parameters

Widget	Type	Description
Text Prompt	Text	Comma-separated object names (e.g., person, car) for detection, or a free-form query for captioning/OCR tasks.
Task	Dropdown	The vision task to execute. See task options below.

Task Options

Task	Description
Object Detection	Detect objects matching the text prompt.
Phrase Grounding	Ground a natural language phrase to image regions.
Open Vocabulary Detection	Detect any concept described in text.
Caption	Generate a short caption for the whole image.
Detailed Caption	Generate a longer, more descriptive caption.
More Detailed Caption	Maximum detail caption.
Dense Region Caption	Caption every detected region individually.
Region Proposal	Return object-like regions without labels.
OCR With Region	Extract text and its bounding box location.

---

Load SAM2 Model Loader

Inputs

None

Widgets

• Model: Select a SAM2 model size.

Outputs

• SAM2 Model: Model object passed to SAM2 Segmentation.

Description

Loads the weights for Segment Anything Model 2 (SAM2). Connect the output to the SAM2 Segmentation node.

Managing Files

• Directorymodels

  • Directorysam2

    • sam2_hiera_tiny.pt
    • sam2_hiera_small.pt
    • sam2_hiera_large.pt

---

SAM2 Segmentation Detector

Inputs

• Image: Input image.
• SAM2 Model: From Load SAM2 Model node.
• points: (Optional) Point prompts.

Widgets

• confidence_threshold: Accuracy filter.

Outputs

• Masks: Segmentation masks.

Description

Segment Anything Model 2 (SAM2). The state-of-the-art model for “cutting out” objects from images. Can be guided with point prompts or run automatically.

Tip

SAM2 is heavy! Use a tiny or small model size for faster processing.

---

Load YOLOX Model Loader

Inputs

None

Widgets

• model_name: Model size selector.

Outputs

• Model: YOLOX Model object.

Description

Loads the weights for the YOLOX object detection architecture. Connect the output to the YOLOX Object Detection node.

Parameters

Widget	Type	Description
model_name	Dropdown	nano (fastest) to x (most accurate).

Managing Files

• Directorymodels

  • Directoryyolox

    • yolox_s.pth
    • yolox_x.pth

---

YOLOX Object Detection Detector

Inputs

• Model: YOLOX Model.
• Image(s): Input Stream.

Widgets

• Confidence: Min score (0.0-1.0).
• NMS: Overlap removal.
• Line/Text Color: Visualization colors.
• Show Labels/Scores: Toggle overlays.

Outputs

• BBoxes: Bounding Boxes.
• Labels: Class names (e.g. “person”).
• Scores: Confidence % (0-1).

Parameters

Widget	Type	Default	Description
confidence_threshold	Float	0.5	Minimum certainty required to detect an object.
nms_threshold	Float	0.45	Non-Maximum Suppression. Higher values merge overlapping boxes less aggressively.
box_color	Color	#00FF00	Color of the bounding box lines.
label_color	Color	#000000	Color of the label text.
show_labels	Bool	True	Draw class names (e.g. “person”) on the image.
show_scores	Bool	True	Draw confidence percentages on the image.

Description

Detects 80 different classes of objects (COCO dataset) in an image, returning their bounding boxes and labels.