`model.predict` fails when running segmentation model on numpy images

import cv2

version = rf.workspace("model-examples").project("car-parts-instance-segmentation").version(3)
dataset = version.download("yolov8")
seg_model = version.model

first_image_path = os.listdir(f"{dataset.location}/train/images")[0]
first_image_path = f"{dataset.location}/train/images/{first_image_path}"
seg_image = cv2.imread(first_image_path)
assert seg_image is not None

# Path
result = seg_model.predict(first_image_path).json()
detections = sv.Detections.from_inference(result)
assert len(detections) > 0
# Succeeds

# Np img
result = seg_model.predict(seg_image).json()
detections = sv.Detections.from_inference(result)
assert len(detections) > 0
# Fails!

# Empty np img
result = seg_model.predict(black_image).json()
detections = sv.Detections.from_inference(result)
assert len(detections) == 0
# Fails!

Start by reproducing the examples in the issue and trace the model.predict entry point for path inputs versus NumPy images. Verify the behavior for the segmentation image and the empty black image; done means both return results that can be converted with sv.Detections.from_inference, with zero detections for the empty image.