SAM 3

Run SAM3 with text prompts for zero-shot segmentation.

v3

Run Segment Anything 3 (SAM3), a zero-shot instance segmentation model, on an image.

You can use text prompts for open-vocabulary segmentation - just specify class names and SAM3 will segment those objects in the image.

This block supports two output formats:

• rle (default): Returns masks in RLE (Run-Length Encoding) format, which is more memory-efficient

• polygons: Returns polygon coordinates for each mask

RLE format is recommended for high-resolution images or workflows with many detections.

Type identifier

Use the following identifier in step "type" field: roboflow_core/sam3@v3 to add the block as a step in your workflow.

Properties

Name	Type	Description	Refs
name	str	Enter a unique identifier for this step..	❌
model_id	str	model version. You only need to change this for fine tuned sam3 models..	✅
class_names	Optional[List[str], str]	List of classes to recognise.	✅
class_mapping	Dict[str, str]	Maps class names in predictions to different output names. Applied after inference, e.g. {'cat': 'gato'} renames 'cat' predictions to 'gato'..	✅
confidence	float	Minimum confidence threshold for predicted masks.	✅
per_class_confidence	List[float]	List of confidence thresholds per class (must match class_names length).	✅
apply_nms	bool	Whether to apply Non-Maximum Suppression across prompts.	✅
nms_iou_threshold	float	IoU threshold for cross-prompt NMS. Must be in [0.0, 1.0].	✅
output_format	str	'rle' returns efficient RLE encoding (recommended), 'polygons' returns polygon coordinates.	❌

The Refs column marks possibility to parametrise the property with dynamic values available in workflow runtime. See Bindings for more info.

Runtime compatibility

hard - runtime self_hosted_cpu; execution local : Requires a GPU; run_locally() loads a model that needs CUDA.

Input and Output Bindings

The available connections depend on its binding kinds. Check what binding kinds SAM 3 in version v3 has.

Input and output bindings

• input

  • images (image): The image to infer on..

  • model_id (roboflow_model_id): model version. You only need to change this for fine tuned sam3 models..

  • class_names (Union[list_of_values, string]): List of classes to recognise.

  • class_mapping (dictionary): Maps class names in predictions to different output names. Applied after inference, e.g. {'cat': 'gato'} renames 'cat' predictions to 'gato'..

  • confidence (float): Minimum confidence threshold for predicted masks.

  • per_class_confidence (list_of_values): List of confidence thresholds per class (must match class_names length).

  • apply_nms (boolean): Whether to apply Non-Maximum Suppression across prompts.

  • nms_iou_threshold (float): IoU threshold for cross-prompt NMS. Must be in [0.0, 1.0].

• output

  • predictions (Union[rle_instance_segmentation_prediction, instance_segmentation_prediction]): Prediction with detected bounding boxes and RLE-encoded segmentation masks in form of sv.Detections(...) object if rle_instance_segmentation_prediction or Prediction with detected bounding boxes and segmentation masks in form of sv.Detections(...) object if instance_segmentation_prediction.

Example JSON definition

{
	    "name": "<your_step_name_here>",
	    "type": "roboflow_core/sam3@v3",
	    "images": "$inputs.image",
	    "model_id": "sam3/sam3_final",
	    "class_names": [
	        "car",
	        "person"
	    ],
	    "class_mapping": {
	        "cat": "gato",
	        "dog": "perro"
	    },
	    "confidence": 0.3,
	    "per_class_confidence": [
	        0.3,
	        0.5,
	        0.7
	    ],
	    "apply_nms": "<block_does_not_provide_example>",
	    "nms_iou_threshold": 0.5,
	    "output_format": "rle"
	}

v2

Run Segment Anything 3, a zero-shot instance segmentation model, on an image.

You can pass in boxes/predictions from other models as prompts, or use a text prompt for open-vocabulary segmentation. If you pass in box detections from another model, the class names of the boxes will be forwarded to the predicted masks.

Type identifier

Use the following identifier in step "type" field: roboflow_core/sam3@v2 to add the block as a step in your workflow.

Properties

Name	Type	Description	Refs
name	str	Enter a unique identifier for this step..	❌
model_id	str	model version. You only need to change this for fine tuned sam3 models..	✅
class_names	Optional[List[str], str]	List of classes to recognise.	✅
confidence	float	Minimum confidence threshold for predicted masks.	✅
per_class_confidence	List[float]	List of confidence thresholds per class (must match class_names length).	✅
apply_nms	bool	Whether to apply Non-Maximum Suppression across prompts.	✅
nms_iou_threshold	float	IoU threshold for cross-prompt NMS. Must be in [0.0, 1.0].	✅

The Refs column marks possibility to parametrise the property with dynamic values available in workflow runtime. See Bindings for more info.

Runtime compatibility

hard - runtime self_hosted_cpu; execution local : Requires a GPU; run_locally() loads a model that needs CUDA.

Input and Output Bindings

The available connections depend on its binding kinds. Check what binding kinds SAM 3 in version v2 has.

Input and output bindings

• input

  • images (image): The image to infer on..

  • model_id (roboflow_model_id): model version. You only need to change this for fine tuned sam3 models..

  • class_names (Union[list_of_values, string]): List of classes to recognise.

  • confidence (float): Minimum confidence threshold for predicted masks.

  • per_class_confidence (list_of_values): List of confidence thresholds per class (must match class_names length).

  • apply_nms (boolean): Whether to apply Non-Maximum Suppression across prompts.

  • nms_iou_threshold (float): IoU threshold for cross-prompt NMS. Must be in [0.0, 1.0].

• output

  • predictions (instance_segmentation_prediction): Prediction with detected bounding boxes and segmentation masks in form of sv.Detections(...) object.

Example JSON definition

{
	    "name": "<your_step_name_here>",
	    "type": "roboflow_core/sam3@v2",
	    "images": "$inputs.image",
	    "model_id": "sam3/sam3_final",
	    "class_names": [
	        "car",
	        "person"
	    ],
	    "confidence": 0.3,
	    "per_class_confidence": [
	        0.3,
	        0.5,
	        0.7
	    ],
	    "apply_nms": "<block_does_not_provide_example>",
	    "nms_iou_threshold": 0.5
	}

v1

Run Segment Anything 3, a zero-shot instance segmentation model, on an image.

You can pass in boxes/predictions from other models as prompts, or use a text prompt for open-vocabulary segmentation. If you pass in box detections from another model, the class names of the boxes will be forwarded to the predicted masks.

Type identifier

Use the following identifier in step "type" field: roboflow_core/sam3@v1 to add the block as a step in your workflow.

Properties

Name	Type	Description	Refs
name	str	Enter a unique identifier for this step..	❌
model_id	str	model version. You only need to change this for fine tuned sam3 models..	✅
class_names	Optional[List[str], str]	List of classes to recognise.	✅
threshold	float	Threshold for predicted mask scores.	✅

The Refs column marks possibility to parametrise the property with dynamic values available in workflow runtime. See Bindings for more info.

Runtime compatibility

hard - runtime self_hosted_cpu; execution local : Requires a GPU; run_locally() loads a model that needs CUDA.

Input and Output Bindings

The available connections depend on its binding kinds. Check what binding kinds SAM 3 in version v1 has.

Input and output bindings

• input

  • images (image): The image to infer on..

  • model_id (roboflow_model_id): model version. You only need to change this for fine tuned sam3 models..

  • class_names (Union[list_of_values, string]): List of classes to recognise.

  • threshold (float): Threshold for predicted mask scores.

• output

  • predictions (instance_segmentation_prediction): Prediction with detected bounding boxes and segmentation masks in form of sv.Detections(...) object.

Example JSON definition

{
	    "name": "<your_step_name_here>",
	    "type": "roboflow_core/sam3@v1",
	    "images": "$inputs.image",
	    "model_id": "sam3/sam3_final",
	    "class_names": [
	        "car",
	        "person"
	    ],
	    "threshold": 0.3
	}