fix(workflows): preserve image dimensions on empty VLM detections

The numpy serialiser returned {"width": null, "height": null} for empty
VLM-as-detector results because sv.Detections.data is per-row — zero rows
means the serialiser per-row loop never executes and IMAGE_DIMENSIONS_KEY
is never read. The tensor-native path stores image dimensions in
image_metadata (a dict on the Detections object), so it correctly emits
real width/height for empty detections.

Fix: store IMAGE_DIMENSIONS_KEY in sv.Detections.metadata (a free-form
dict that survives zero rows) and teach the numpy serialiser to read it as
a fallback when the per-row loop yields no rows.

Before: serialise_sv_detections(parse_...(empty_input)) returns
  {image: {width: null, height: null}, predictions: []}

After: serialise_sv_detections(parse_...(empty_input)) returns
  {image: {width: 640, height: 480}, predictions: []}

All 7 VLM-as-detector parsers updated (anthropic, gemini, muse, openai,
qwen, spacexai, v2/llm). 1905 tests pass.

Author: davidnichols-ops

inference/core/workflows/core_steps/common/serializers.py

@@ -250,10 +250,16 @@ def serialise_sv_detections(detections: sv.Detections) -> dict:
     }  # TODO: this breaks the contract of
     # standard inference, but to fix that problem, we would need sv.Detections to provide
     # detection-level metadata.
+    if image_dimensions is None:
+        # Zero-row detections carry image dimensions in ``metadata`` (not per-row
+        # ``data``, which the loop above never iterates). Producers that return
+        # empty detections with ``metadata={IMAGE_DIMENSIONS_KEY: [h, w]}`` get
+        # the same ``image.width`` / ``image.height`` as the tensor-native path.
+        image_dimensions = detections.metadata.get(IMAGE_DIMENSIONS_KEY)
     if image_dimensions is not None:
         image_metadata = {
-            "width": image_dimensions[1].item(),
-            "height": image_dimensions[0].item(),
+            "width": int(image_dimensions[1]),
+            "height": int(image_dimensions[0]),
         }
     return {"image": image_metadata, "predictions": serialized_detections}
 

inference/core/workflows/core_steps/common/utils.py

@@ -228,6 +228,26 @@ def attach_parents_coordinates_to_batch_of_sv_detections(
     return result
 
 
+def empty_detections_with_image_metadata(
+    image_height: int,
+    image_width: int,
+) -> sv.Detections:
+    """Create a zero-row ``sv.Detections`` that carries image dimensions in
+    ``metadata`` so the numpy serialiser can emit real ``image.width`` /
+    ``image.height`` for empty VLM results — matching the tensor-native path.
+
+    ``sv.Detections.data`` is per-row, so zero rows means the serialiser's
+    per-row loop never executes and any keys stored there are invisible.
+    ``metadata`` is a free-form dict on ``sv.Detections`` that survives zero
+    rows and is read by the serialiser as a fallback when no per-row
+    ``IMAGE_DIMENSIONS_KEY`` is found.
+    """
+    return sv.Detections(
+        xyxy=np.empty((0, 4), dtype=np.float32),
+        metadata={IMAGE_DIMENSIONS_KEY: [image_height, image_width]},
+    )
+
+
 def attach_parents_coordinates_to_sv_detections(
     detections: sv.Detections,
     image: WorkflowImageData,

inference/core/workflows/core_steps/formatters/vlm_as_detector/anthropic_detection_parsing.py

@@ -8,6 +8,7 @@
 from inference.core.workflows.core_steps.common.utils import (
     attach_parents_coordinates_to_sv_detections,
     compute_anthropic_upload_dimensions,
+    empty_detections_with_image_metadata,
 )
 from inference.core.workflows.core_steps.formatters.vlm_as_detector.gemini_detection_parsing import (
     create_classes_index,
@@ -88,7 +89,11 @@ def parse_anthropic_object_detection_response(
     if not isinstance(parsed_data, list):
         raise ValueError("Unexpected Anthropic Claude object detection response format")
     if len(parsed_data) == 0:
-        return sv.Detections.empty()
+        image_height, image_width = image.numpy_image.shape[:2]
+        return empty_detections_with_image_metadata(
+            image_height=image_height,
+            image_width=image_width,
+        )
 
     class_name2id = create_classes_index(classes=classes)
     image_height, image_width = image.numpy_image.shape[:2]

inference/core/workflows/core_steps/formatters/vlm_as_detector/gemini_detection_parsing.py

@@ -7,6 +7,7 @@
 
 from inference.core.workflows.core_steps.common.utils import (
     attach_parents_coordinates_to_sv_detections,
+    empty_detections_with_image_metadata,
 )
 from inference.core.workflows.execution_engine.constants import (
     DETECTION_ID_KEY,
@@ -77,7 +78,10 @@ def parse_gemini_object_detection_response(
     image_height, image_width = image.numpy_image.shape[:2]
     detections = extract_gemini_detection_entries(parsed_data=parsed_data)
     if len(detections) == 0:
-        return sv.Detections.empty()
+        return empty_detections_with_image_metadata(
+            image_height=image_height,
+            image_width=image_width,
+        )
 
     xyxy, class_id, class_name, confidence = [], [], [], []
     for detection in detections:

inference/core/workflows/core_steps/formatters/vlm_as_detector/muse_detection_parsing.py

@@ -10,6 +10,7 @@
 from inference.core.logger import logger
 from inference.core.workflows.core_steps.common.utils import (
     attach_parents_coordinates_to_sv_detections,
+    empty_detections_with_image_metadata,
 )
 from inference.core.workflows.core_steps.formatters.vlm_as_detector.gemini_detection_parsing import (
     create_classes_index,
@@ -195,7 +196,10 @@ def parse_muse_object_detection_response(
         confidence.append(1.0)
 
     if not xyxy:
-        return sv.Detections.empty()
+        return empty_detections_with_image_metadata(
+            image_height=image_height,
+            image_width=image_width,
+        )
 
     xyxy = np.array(xyxy).round(0)
     detection_ids = np.array([str(uuid4()) for _ in range(len(xyxy))])

inference/core/workflows/core_steps/formatters/vlm_as_detector/openai_detection_parsing.py

@@ -8,6 +8,7 @@
 from inference.core.workflows.core_steps.common.utils import (
     DETECTION_MAX_EDGE_PIXELS,
     attach_parents_coordinates_to_sv_detections,
+    empty_detections_with_image_metadata,
     scale_dimensions_to_max_edge,
 )
 from inference.core.workflows.core_steps.formatters.vlm_as_detector.gemini_detection_parsing import (
@@ -89,7 +90,11 @@ def parse_openai_object_detection_response(
     if not isinstance(parsed_data, list):
         raise ValueError("Unexpected OpenAI object detection response format")
     if len(parsed_data) == 0:
-        return sv.Detections.empty()
+        image_height, image_width = image.numpy_image.shape[:2]
+        return empty_detections_with_image_metadata(
+            image_height=image_height,
+            image_width=image_width,
+        )
 
     class_name2id = create_classes_index(classes=classes)
     image_height, image_width = image.numpy_image.shape[:2]

inference/core/workflows/core_steps/formatters/vlm_as_detector/qwen_detection_parsing.py

@@ -8,6 +8,7 @@
 from inference.core.logger import logger
 from inference.core.workflows.core_steps.common.utils import (
     attach_parents_coordinates_to_sv_detections,
+    empty_detections_with_image_metadata,
 )
 from inference.core.workflows.core_steps.formatters.vlm_as_detector.gemini_detection_parsing import (
     create_classes_index,
@@ -183,7 +184,10 @@ def parse_qwen_object_detection_response(
         confidence.append(1.0)
 
     if not xyxy:
-        return sv.Detections.empty()
+        return empty_detections_with_image_metadata(
+            image_height=image_height,
+            image_width=image_width,
+        )
 
     xyxy = np.array(xyxy).round(0)
     confidence = np.array(confidence)

inference/core/workflows/core_steps/formatters/vlm_as_detector/spacexai_detection_parsing.py

@@ -7,6 +7,7 @@
 
 from inference.core.workflows.core_steps.common.utils import (
     attach_parents_coordinates_to_sv_detections,
+    empty_detections_with_image_metadata,
 )
 from inference.core.workflows.core_steps.formatters.vlm_as_detector.gemini_detection_parsing import (
     create_classes_index,
@@ -99,7 +100,11 @@ def parse_spacexai_object_detection_response(
     """
     detections = extract_spacexai_detection_entries(parsed_data=parsed_data)
     if len(detections) == 0:
-        return sv.Detections.empty()
+        image_height, image_width = image.numpy_image.shape[:2]
+        return empty_detections_with_image_metadata(
+            image_height=image_height,
+            image_width=image_width,
+        )
 
     class_name2id = create_classes_index(classes=classes)
     image_height, image_width = image.numpy_image.shape[:2]

inference/core/workflows/core_steps/formatters/vlm_as_detector/v2.py

@@ -13,6 +13,7 @@
 
 from inference.core.workflows.core_steps.common.utils import (
     attach_parents_coordinates_to_sv_detections,
+    empty_detections_with_image_metadata,
 )
 from inference.core.workflows.core_steps.common.vlms import VLM_TASKS_METADATA
 from inference.core.workflows.core_steps.formatters.vlm_as_detector.anthropic_detection_parsing import (
@@ -398,7 +399,10 @@ def parse_llm_object_detection_response(
     class_name2id = create_classes_index(classes=classes)
     image_height, image_width = image.numpy_image.shape[:2]
     if len(parsed_data["detections"]) == 0:
-        return sv.Detections.empty()
+        return empty_detections_with_image_metadata(
+            image_height=image_height,
+            image_width=image_width,
+        )
     xyxy, class_id, class_name, confidence = [], [], [], []
     for detection in parsed_data["detections"]:
         xyxy.append(

tests/workflows/unit_tests/core_steps/common/test_serializers.py

@@ -195,6 +195,34 @@ def test_serialise_sv_detections() -> None:
     }
 
 
+def test_serialise_sv_detections_empty_without_metadata_returns_none_dims() -> None:
+    # given — plain sv.Detections.empty() has no metadata, so the serialiser
+    # cannot recover image dimensions (the historical behaviour).
+    detections = sv.Detections.empty()
+
+    # when
+    result = serialise_sv_detections(detections=detections)
+
+    # then
+    assert result == {"image": {"width": None, "height": None}, "predictions": []}
+
+
+def test_serialise_sv_detections_empty_with_metadata_returns_real_dims() -> None:
+    # given — zero-row detections carrying IMAGE_DIMENSIONS_KEY in metadata
+    # (the contract producers use to pass image-level info past zero rows).
+    detections = sv.Detections(
+        xyxy=np.empty((0, 4), dtype=np.float32),
+        metadata={"image_dimensions": [480, 640]},
+    )
+
+    # when
+    result = serialise_sv_detections(detections=detections)
+
+    # then — the serialiser reads image dimensions from metadata as a fallback
+    # when the per-row loop never executes, matching the tensor-native path.
+    assert result == {"image": {"width": 640, "height": 480}, "predictions": []}
+
+
 def test_serialise_sv_detections_skips_padded_keypoint_slots() -> None:
     # given the padded, rectangular (n, max_kps, 2) keypoint layout that
     # add_inference_keypoints_to_sv_detections produces when detections carry