diff --git a/src/poseforge/pose/camera.py b/src/poseforge/pose/camera.py
index b7718ed3..89d83744 100644
--- a/src/poseforge/pose/camera.py
+++ b/src/poseforge/pose/camera.py
@@ -95,8 +95,17 @@ def __init__(
     def __call__(self, xy, depth):
         """Map 2D camera coordinates + depth to 3D world coordinates.
 
+        IMPORTANT: ``xy`` MUST be expressed in the SAME pixel space the mapper
+        was built with, i.e. in ``rendering_size`` pixels. The focal length and
+        principal point scale linearly with ``rendering_size``, so feeding xy
+        that live in a different pixel resolution injects an anisotropic
+        distortion (in-plane x, y scaled by the resolution ratio, depth left
+        exact). If your predictions are at resolution S, build the mapper with
+        ``rendering_size=(S, S)``.
+
         Args:
-            xy: (..., 2) array of 2D camera coordinates (in pixels).
+            xy: (..., 2) array of 2D camera coordinates, in ``rendering_size``
+                pixels.
             depth: (...,) array of depth values (in mm).
 
         Returns:
diff --git a/src/poseforge/pose/keypoints3d/scripts/run_keypoints3d_inference.py b/src/poseforge/pose/keypoints3d/scripts/run_keypoints3d_inference.py
index f59d16af..d55755d7 100644
--- a/src/poseforge/pose/keypoints3d/scripts/run_keypoints3d_inference.py
+++ b/src/poseforge/pose/keypoints3d/scripts/run_keypoints3d_inference.py
@@ -97,9 +97,32 @@ def run_keypoints3d_inference(
     print("========== Model Summary ==========")
     summary(pipeline.model, input_size=(3, *inference_image_size), device="cuda")
 
-    # Set up camera mapper
+    # Set up camera mapper.
+    #
+    # IMPORTANT: the mapper's intrinsics (focal length and principal point)
+    # scale linearly with the image/sensor size it is built with, so it MUST be
+    # built at the pixel space the predictions actually live in. The model
+    # consumes images resized to `inference_image_size` and emits `pred_xy` in
+    # that same pixel space. Therefore the mapper is built at
+    # `inference_image_size`, NOT at `camera_rendering_size`.
+    #
+    # `camera_rendering_size` is the resolution at which the *training* data was
+    # rendered in simulation; it is NOT the space the predictions are in. Using
+    # it here would scale the in-plane (x, y) coordinates by
+    # inference_image_size / camera_rendering_size while leaving depth exact,
+    # injecting an anisotropic distortion that corrupts downstream IK joint
+    # angles. See tests/test_camera_unprojection.py.
+    if tuple(camera_rendering_size) != tuple(inference_image_size):
+        print(
+            "Note: camera_rendering_size "
+            f"{tuple(camera_rendering_size)} differs from inference_image_size "
+            f"{tuple(inference_image_size)}. The camera mapper is built at "
+            "inference_image_size because pred_xy is in inference_image_size "
+            "pixels; camera_rendering_size (the simulation render size) is "
+            "intentionally not used for unprojection."
+        )
     cam_mapper = CameraToWorldMapper(
-        camera_pos, camera_fov_deg, camera_rendering_size, camera_rotation_euler
+        camera_pos, camera_fov_deg, inference_image_size, camera_rotation_euler
     )
 
     # Run inference
@@ -287,7 +310,7 @@ def start():
         output_basedir=output_basedir,
         batch_size=batch_size,
         n_workers=n_workers,
-        c=inference_image_size,
+        inference_image_size=inference_image_size,
         camera_pos=camera_pos,
         camera_fov_deg=camera_fov_deg,
         camera_rendering_size=camera_rendering_size,
diff --git a/src/poseforge/production/spotlight/keypoints3d.py b/src/poseforge/production/spotlight/keypoints3d.py
index 544a006e..dc90187e 100644
--- a/src/poseforge/production/spotlight/keypoints3d.py
+++ b/src/poseforge/production/spotlight/keypoints3d.py
@@ -62,14 +62,37 @@ def predict_keypoints3d(
     logger.info("Creating 3D keypoints inference pipeline")
     pipeline = keypoints3d.Pose2p5DPipeline(model, device=device, use_float16=True)
 
-    # Set up camera mapper
+    # Set up camera mapper.
+    #
+    # IMPORTANT: the mapper's intrinsics (focal length and principal point)
+    # scale linearly with the image/sensor size it is built with, so it MUST be
+    # built at the pixel space the predictions actually live in. The model
+    # consumes images resized to `working_size` and emits `pred_xy` in that same
+    # pixel space, so the mapper is built at `(working_size, working_size)`, NOT
+    # at `camera_rendering_size`. `camera_rendering_size` is the simulation
+    # render resolution of the training data and is NOT the prediction space;
+    # using it would scale in-plane (x, y) by working_size / camera_rendering_size
+    # while leaving depth exact, an anisotropic distortion that corrupts
+    # downstream IK joint angles. See tests/test_camera_unprojection.py.
+    working_size = keypoints3d_model_config["working_size"]
+    if tuple(camera_rendering_size) != (working_size, working_size):
+        logger.info(
+            "camera_rendering_size {} differs from working_size {}; building the "
+            "camera mapper at working_size because pred_xy is in working_size "
+            "pixels (camera_rendering_size is the simulation render size and is "
+            "intentionally not used for unprojection).",
+            tuple(camera_rendering_size),
+            (working_size, working_size),
+        )
     cam_mapper = CameraToWorldMapper(
-        camera_pos, camera_fov_deg, camera_rendering_size, camera_rotation_euler
+        camera_pos,
+        camera_fov_deg,
+        (working_size, working_size),
+        camera_rotation_euler,
     )
 
     # Create video loader
     logger.info("Creating video loader for 3D keypoints prediction")
-    working_size = keypoints3d_model_config["working_size"]
     video_loader = SimpleVideoCollectionLoader(
         [aligned_behavior_video_path],
         transform=transforms.Resize((working_size, working_size)),
diff --git a/tests/test_camera_unprojection.py b/tests/test_camera_unprojection.py
new file mode 100644
index 00000000..2293525b
--- /dev/null
+++ b/tests/test_camera_unprojection.py
@@ -0,0 +1,199 @@
+"""Tests for ``poseforge.pose.camera.CameraToWorldMapper``.
+
+These tests pin down two facts that motivate the bug fix in
+``run_keypoints3d_inference.py`` and ``production/spotlight/keypoints3d.py``:
+
+1. The mapper is a correct inverse of its own forward projection (round-trip
+   recovery to ~1e-6).
+2. The mapper's intrinsics scale with the pixel resolution it is built with:
+   the in-plane (x, y) focal scaling is proportional to the image size while
+   depth is left untouched. Consequently, unprojecting predictions that live in
+   ``inference_image_size`` (=256) pixels with a mapper built at 256 is correct,
+   whereas building it at the simulation render size (=464) injects an
+   anisotropic distortion (in-plane factor 256/464 ~= 0.55, depth factor 1.0).
+
+``camera.py`` depends only on numpy and scipy, so these tests run standalone.
+The module is loaded directly from the source file so the tests always exercise
+the code under review regardless of where an editable install resolves
+``poseforge`` to.
+"""
+
+import importlib.util
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+# Load camera.py directly from this repo's source tree (sibling of tests/).
+_CAMERA_PY = (
+    Path(__file__).resolve().parent.parent
+    / "src"
+    / "poseforge"
+    / "pose"
+    / "camera.py"
+)
+_spec = importlib.util.spec_from_file_location("_poseforge_camera_under_test", _CAMERA_PY)
+_camera = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(_camera)
+CameraToWorldMapper = _camera.CameraToWorldMapper
+
+
+# Camera parameters matching the inference/production defaults.
+CAMERA_POS = (0.0, 0.0, -100.0)
+CAMERA_FOV_DEG = 5.0
+CAMERA_ROTATION_EULER = (0.0, np.pi, -np.pi / 2)
+
+INFERENCE_SIZE = 256  # pixel space the model's pred_xy actually lives in
+RENDERING_SIZE = 464  # simulation render size -- NOT the prediction space
+
+
+def _make_mapper(size: int) -> "CameraToWorldMapper":
+    return CameraToWorldMapper(
+        CAMERA_POS, CAMERA_FOV_DEG, (size, size), CAMERA_ROTATION_EULER
+    )
+
+
+def _forward_project(mapper, world_xyz: np.ndarray):
+    """Project world points to (pixel_xy, depth) with the mapper's own
+    ``camera_matrix``. This is the exact inverse operation of ``__call__``.
+
+    Args:
+        world_xyz: (N, 3) world coordinates in mm.
+
+    Returns:
+        (pixel_xy (N, 2), depth (N,)).
+    """
+    hom = np.hstack([world_xyz, np.ones((world_xyz.shape[0], 1))])  # (N, 4)
+    proj = (mapper.camera_matrix @ hom.T).T  # (N, 3): [x*z, y*z, z]
+    depth = proj[:, 2]
+    pixel_xy = proj[:, :2] / proj[:, 2:3]
+    return pixel_xy, depth
+
+
+def _to_camera_frame(mapper, world_xyz: np.ndarray) -> np.ndarray:
+    """Map world coords into the mapper's camera frame. The camera-frame z is
+    the depth; x, y are the in-plane coordinates."""
+    hom = np.hstack([world_xyz, np.ones((world_xyz.shape[0], 1))])
+    return (mapper.camera_extrinsic_mat @ hom.T).T[:, :3]
+
+
+@pytest.mark.parametrize("size", [128, 256, 464])
+def test_unprojection_inverts_projection(size):
+    """Unprojecting the mapper's own forward projection recovers the original
+    world points to numerical precision."""
+    mapper = _make_mapper(size)
+
+    # Several known world points spread around the camera target region (mm).
+    world_xyz = np.array(
+        [
+            [0.5, -1.0, 0.2],
+            [1.0, -2.0, 0.5],
+            [-0.3, 0.4, -0.1],
+            [2.0, 1.5, 1.0],
+            [-1.2, -0.7, 0.9],
+            [0.0, 0.0, 0.0],
+        ]
+    )
+
+    pixel_xy, depth = _forward_project(mapper, world_xyz)
+    recovered = mapper(pixel_xy, depth)
+
+    assert recovered.shape == world_xyz.shape
+    np.testing.assert_allclose(recovered, world_xyz, atol=1e-6, rtol=0)
+
+
+def test_focal_scales_with_size_depth_does_not():
+    """The in-plane focal scaling is linear in image size; depth is independent
+    of image size.
+
+    This is the root cause of the bug: the mapper's geometry depends on the
+    pixel space it is built with, so it must be built at the resolution the
+    predictions live in.
+    """
+    m256 = _make_mapper(INFERENCE_SIZE)
+    m464 = _make_mapper(RENDERING_SIZE)
+
+    focal_256 = -m256.focal_transform_mat[0, 0]
+    focal_464 = -m464.focal_transform_mat[0, 0]
+
+    # In-plane focal scaling is proportional to image size.
+    np.testing.assert_allclose(
+        focal_256 / focal_464, INFERENCE_SIZE / RENDERING_SIZE, rtol=1e-12
+    )
+    # ~0.55, NOT 1.0 -- so resolution matters for in-plane coords.
+    assert abs(focal_256 / focal_464 - 0.55) < 0.01
+
+    # The depth axis of the focal transform is the identity regardless of size.
+    assert m256.focal_transform_mat[2, 2] == 1.0
+    assert m464.focal_transform_mat[2, 2] == 1.0
+
+
+def test_same_input_different_mapper_size_anisotropic_distortion():
+    """Unprojecting the SAME (256-space) xy + depth with a 256-mapper vs a
+    464-mapper differs, with in-plane scaling ~= 256/464 and depth scaling
+    exactly 1.0.
+
+    This confirms that 256 is the correct resolution for 256-space predictions:
+    using a 464-mapper shrinks the recovered in-plane geometry by ~0.55 while
+    leaving depth untouched -- precisely the anisotropy that corrupts the
+    downstream inverse-kinematics joint angles.
+    """
+    m256 = _make_mapper(INFERENCE_SIZE)  # correct for 256-space pred_xy
+    m464 = _make_mapper(RENDERING_SIZE)  # buggy: 256-space xy fed to a 464 mapper
+
+    # pred_xy lives in 256-px space; depths are mm.
+    xy = np.array(
+        [
+            [100.0, 120.0],
+            [200.0, 50.0],
+            [10.0, 240.0],
+            [128.0, 128.0],
+            [60.0, 200.0],
+        ]
+    )
+    depth = np.array([95.0, 102.0, 100.0, 99.0, 98.0])
+
+    world_256 = m256(xy, depth)
+    world_464 = m464(xy, depth)
+
+    # The two unprojections genuinely disagree (this is the bug's effect).
+    assert not np.allclose(world_256, world_464)
+
+    cam_256 = _to_camera_frame(m256, world_256)
+    cam_464 = _to_camera_frame(m464, world_464)
+
+    # Depth (camera-frame z) is exactly the input depth in BOTH cases: depth is
+    # unaffected by the mapper resolution -> depth ratio is exactly 1.0.
+    np.testing.assert_allclose(cam_256[:, 2], depth, atol=1e-9)
+    np.testing.assert_allclose(cam_464[:, 2], depth, atol=1e-9)
+    np.testing.assert_allclose(cam_256[:, 2] / cam_464[:, 2], 1.0, atol=1e-9)
+
+    # In-plane geometry, however, is scaled. Hold the pixel displacement away
+    # from each sensor's own principal point fixed (this isolates 1/focal): the
+    # same +(dx, dy) pixel step maps to a camera-frame in-plane offset whose
+    # magnitude scales as focal_464 / focal_256 = 464/256. Equivalently, the
+    # 256-mapper's in-plane extent is 256/464 ~= 0.55x that of the 464-mapper
+    # for the same predicted pixels.
+    fixed_depth = np.array([100.0, 100.0])
+    step = np.array([40.0, 25.0])  # arbitrary fixed pixel displacement
+
+    def inplane_step_magnitude(mapper):
+        pp = (mapper.rendering_size[0] - 1) / 2.0
+        base_px = np.array([pp, pp])
+        query_xy = np.stack([base_px, base_px + step])
+        cam = _to_camera_frame(mapper, mapper(query_xy, fixed_depth))
+        return np.linalg.norm(cam[1, :2] - cam[0, :2])
+
+    inplane_256 = inplane_step_magnitude(m256)
+    inplane_464 = inplane_step_magnitude(m464)
+
+    # In-plane displacement ratio is the inverse of the focal ratio.
+    np.testing.assert_allclose(
+        inplane_256 / inplane_464, RENDERING_SIZE / INFERENCE_SIZE, rtol=1e-9
+    )
+    # Equivalently: the 464-mapper's in-plane response is 256/464 ~= 0.55x the
+    # 256-mapper's, while depth response is identical (1.0x). Anisotropy.
+    np.testing.assert_allclose(
+        inplane_464 / inplane_256, INFERENCE_SIZE / RENDERING_SIZE, rtol=1e-9
+    )
+    assert abs(inplane_464 / inplane_256 - 0.55) < 0.01