@@ -192,6 +192,19 @@ class ProjectConfig(DLCVersionedConfig):
     )
     croppedtraining: bool | None = None
 
+    # TODO @deruyter92 2026-09-01: This should be included as field in the
+    # config. It requires a version bump of the schema version.
+    # see https://github.com/DeepLabCut/DeepLabCut/pull/3470
+    @property
+    def bboxes_pcutoff(self) -> Fraction:
+        """Default confidence cutoff for plotting bounding boxes from detectors.
+
+        Note: This property is introduced in v3.1.0 as prospective config
+        field, but is intentionally not yet defined as such in the schema to
+        avoid breaking compatibility with older configs.
+        """
+        return 0.6
+
     @property
     def bodyparts_list(self) -> list[str]:
         # Animal-count agnostic; Always return a list (never "MULTI!", None, etc.)
 
@@ -462,6 +462,7 @@ def evaluate_snapshot(
     per_keypoint_evaluation: bool = False,
     detector_snapshot: Snapshot | None = None,
     pcutoff: float | list[float] | dict[str, float] | None = None,
+    bboxes_pcutoff: float | None = None,
 ) -> pd.DataFrame:
     """Evaluates a snapshot. The evaluation results are stored in the .h5 and .csv file
     under the subdirectory 'evaluation_results'.
@@ -490,7 +491,12 @@ def evaluate_snapshot(
             (if there are any). If a dict is provided, the keys should be bodyparts
             mapping to pcutoff values for each bodypart. Bodyparts that are not defined
             in the dict will have pcutoff set to 0.6.
+        bboxes_pcutoff: The cutoff to use for plotting bounding boxes from detectors.
+            When `None`, the cutoff will be loaded from the project config.
     """
+    if bboxes_pcutoff is None:
+        bboxes_pcutoff = cfg.bboxes_pcutoff
+
     head_type = loader.model_cfg["model"]["heads"]["bodypart"]["type"]
     if head_type == "DLCRNetHead":
         prune_paf_graph.benchmark_paf_graphs(
@@ -622,8 +628,6 @@ def evaluate_snapshot(
 
         df_ground_truth = ensure_multianimal_df_format(loader.df)
 
-        bboxes_cutoff = loader.model_cfg.select("detector.model.box_score_thresh") or 0.6
-
         for mode in ["train", "test"]:
             df_combined = predictions[mode].merge(df_ground_truth, left_index=True, right_index=True)
             bboxes_split = bounding_boxes[mode]
@@ -642,7 +646,7 @@ def evaluate_snapshot(
                 alpha_value=cfg["alphavalue"],
                 p_cutoff=cfg["pcutoff"],
                 bounding_boxes=bboxes_split,
-                bboxes_cutoff=bboxes_cutoff,
+                bboxes_cutoff=bboxes_pcutoff,
             )
 
     return df_predictions
 
@@ -10,7 +10,7 @@
 #
 """Model configuration class for DeepLabCut pose estimation models."""
 
-from pydantic import Field
+from pydantic import Field, field_validator
 
 from deeplabcut.core.config import DLCBaseConfig
 from deeplabcut.core.config.validation import Fraction
@@ -42,10 +42,20 @@ class DetectorModelConfig(DLCBaseConfig):
         freeze_bn_stats: Whether to freeze batch normalization statistics
         freeze_bn_weights: Whether to freeze batch normalization weights
         variant: Specific variant of the detector model
+        box_score_thresh: The score below which the detector discards proposals
+            internally. ``null`` is normalized to the default. The value is
+            deliberately permissive, as raising this value silently drops
+            detections. Note the distinction with ``bboxes_pcutoff`` in the
+            project config, which is the cutoff used to *plot* bounding boxes.
     """
 
     type: str = ""
     freeze_bn_stats: bool = False
     freeze_bn_weights: bool = False
     variant: str | None = None
-    box_score_thresh: Fraction | None = None
+    box_score_thresh: Fraction = 0.01
+
+    @field_validator("box_score_thresh", mode="before")
+    @classmethod
+    def normalize_box_score_thresh(cls, v):
+        return cls.model_fields["box_score_thresh"].default if v is None else v
@@ -510,8 +510,8 @@ def create_labeled_video(
             When True, f(x) = max(0, (x - pcutoff)/(1 - pcutoff)). Defaults to False.
         plot_bboxes (bool, optional): If using Pytorch and in Top-Down mode,
             setting this to true will also plot the bounding boxes. Defaults to True.
-        bboxes_pcutoff (float, optional): If plotting bounding boxes, this overrides the bboxes_pcutoff
-            set in the model configuration. Defaults to None.
+        bboxes_pcutoff (float, optional): Minimum bounding box confidence required for plotting.
+            When ``None``, the value is loaded from the project configuration. Defaults to None.
         max_workers (int | None): Maximum number of processes to use for multiprocessing.
             Set this parameter to limit the total RAM-usage of simultaneous processes.
             Default: no maximum (i.e. number of spawned processes is based on the number of
@@ -585,6 +585,8 @@ def create_labeled_video(
         track_method = auxfun_multianimal.get_track_method(cfg, track_method=track_method)
         if pcutoff is None:
             pcutoff = cfg["pcutoff"]
+        if bboxes_pcutoff is None:
+            bboxes_pcutoff = cfg.bboxes_pcutoff
 
         # Get individuals from the config
         individuals = cfg.get("individuals", [""])
@@ -604,11 +606,6 @@ def create_labeled_video(
             model_config = PoseConfig.from_yaml(model_config_path)
             if model_config.select("train_settings.weight_init.memory_replay"):
                 superanimal_name = model_config["train_settings"]["weight_init"]["dataset"]
-            if bboxes_pcutoff is None:
-                bboxes_pcutoff = model_config.select("detector.model.box_score_thresh") or 0.6
-        else:
-            if bboxes_pcutoff is None:
-                bboxes_pcutoff = 0.6
 
     if init_weights == "":
         DLCscorer, DLCscorerlegacy = auxiliaryfunctions.get_scorer_name(
@@ -1084,6 +1081,7 @@ def create_video_with_all_detections(
     modelprefix="",
     confidence_to_alpha: bool | Callable[[float], float] = False,
     plot_bboxes: bool = True,
+    bboxes_pcutoff: float | None = None,
     **kwargs,
 ):
     """Create a video labeled with all the detections stored in a '*_full.pickle' file.
@@ -1118,6 +1116,8 @@ def create_video_with_all_detections(
         plot_bboxes (bool, optional): If detections were produced using a Pytorch Top-Down model,
             setting this parameter to True will also plot
             the bounding boxes generated by the detector. Defaults to True.
+        bboxes_pcutoff (float, optional): Minimum bounding box confidence required for plotting.
+            When ``None``, the value is loaded from the project configuration. Defaults to None.
         kwargs (dict, optional): Additional arguments.
             For torch-based shuffles, can be used to specify:
                 - snapshot_index
@@ -1127,6 +1127,7 @@ def create_video_with_all_detections(
 
     from deeplabcut.core.inferenceutils import Assembler
 
+    # TODO @deruyter92 2026-09-01: This should be refactored to ProjectConfig.from_any()
     cfg = auxiliaryfunctions.read_config(config)
     trainFraction = cfg["TrainingFraction"][trainingsetindex]
     DLCscorername, _ = auxiliaryfunctions.get_scorer_name(
@@ -1137,6 +1138,9 @@ def create_video_with_all_detections(
         **kwargs,
     )
 
+    if bboxes_pcutoff is None:
+        bboxes_pcutoff = cfg.bboxes_pcutoff
+
     videos = collect_video_paths(videos, extensions=video_extensions)
     if not videos:
         return
@@ -1193,9 +1197,6 @@ def create_video_with_all_detections(
             clip = vp(fname=video, sname=outputname, codec="mp4v")
             ny, nx = clip.height, clip.width
 
-            bboxes_pcutoff = 0.6
-            if pytorch_cfg := (metadata.get("data") or {}).get("pytorch-config"):
-                bboxes_pcutoff = PoseConfig.from_any(pytorch_cfg).select("detector.model.box_score_thresh") or 0.6
             bboxes_color = (255, 0, 0)
 
             for n in trange(clip.nframes):
 
@@ -309,3 +309,26 @@ def test_unversioned_yaml_treated_as_v0(self, tmp_path):
         config_path.write_text(f"project_path: {tmp_path}\nengine: pytorch\n")
         cfg = ProjectConfig.from_yaml(config_path)
         assert cfg.config_version == CURRENT_CONFIG_VERSION
+
+
+# -----------------------------------------------------------------------------
+# Fields awaiting the next schema version bump
+#
+# Each test here describes a field that is exposed as a read-only property today
+# because promoting it to a real schema field requires bumping
+# CURRENT_CONFIG_VERSION and writing the migration. They are marked strict xfail
+# so that the migration landing turns them into failures, as a reminder to drop
+# the property and the marker together.
+# -----------------------------------------------------------------------------
+
+
+class TestFieldsPendingSchemaMigration:
+    @pytest.mark.xfail(
+        strict=True,
+        reason=(
+            "bboxes_pcutoff is a property until the schema version is bumped, "
+            "see https://github.com/DeepLabCut/DeepLabCut/pull/3470"
+        ),
+    )
+    def test_bboxes_pcutoff_is_a_schema_field(self):
+        assert "bboxes_pcutoff" in ProjectConfig.model_fields
@@ -167,7 +167,7 @@ def test_pose_config_creation(case: PoseConfigCase, tmp_path: Path) -> None:
         assert detector is not None
         assert detector["model"]["type"] == case.detector_type
         assert detector["device"] == "auto"
-        assert detector["model"].get("box_score_thresh") is None
+        assert detector["model"]["box_score_thresh"] == 0.01
         assert cfg["data"]["train"].get("top_down_crop") is not None
     else:
         assert cfg.get("detector") is None
 
@@ -0,0 +1,58 @@
+#
+# DeepLabCut Toolbox (deeplabcut.org)
+# © A. & M.W. Mathis Labs
+# https://github.com/DeepLabCut/DeepLabCut
+#
+# Please see AUTHORS for contributors.
+# https://github.com/DeepLabCut/DeepLabCut/blob/main/AUTHORS
+#
+# Licensed under GNU Lesser General Public License v3.0
+#
+"""Generated detector configs store ``box_score_thresh: null``; building a
+Faster R-CNN from such a config must not leave torchvision's score threshold
+as None, which crashes ``postprocess_detections`` during inference."""
+
+import pytest
+
+from deeplabcut.pose_estimation_pytorch.config.model import DetectorModelConfig
+from deeplabcut.pose_estimation_pytorch.models.detectors.base import DETECTORS
+
+
+@pytest.mark.parametrize(
+    ("value", "expected"),
+    [
+        (None, 0.01),
+        (0.0, 0.0),
+        (0.4, 0.4),
+    ],
+)
+def test_detector_box_score_thresh(value, expected):
+    config = DetectorModelConfig(box_score_thresh=value)
+
+    assert config.box_score_thresh == expected
+
+
+def test_detector_box_score_thresh_default():
+    config = DetectorModelConfig()
+
+    assert config.box_score_thresh == 0.01
+
+
+def test_build_from_config_with_null_threshold():
+    """The detector must not be built with torchvision's score_thresh set to None.
+
+    Nothing at the model layer normalizes the threshold any more, so this asserts
+    the property the config layer is relied upon for: a legacy ``null`` never
+    reaches the torchvision constructor.
+    """
+    config = DetectorModelConfig.from_dict(
+        {
+            "type": "FasterRCNN",
+            "variant": "fasterrcnn_mobilenet_v3_large_fpn",
+            "box_score_thresh": None,
+        }
+    )
+
+    detector = DETECTORS.build(config)
+
+    assert detector.model.roi_heads.score_thresh == 0.01