diff --git a/tests/test_focus_estimator.py b/tests/test_focus_estimator.py
index 5e4442df..8c093195 100644
--- a/tests/test_focus_estimator.py
+++ b/tests/test_focus_estimator.py
@@ -1,5 +1,6 @@
 import numpy as np
 import pytest
+import torch
 
 from waveorder import focus
 
@@ -85,3 +86,73 @@ def test_focus_estimator_snr(tmp_path):
         assert plot_path.exists()
         if slice is not None:
             assert np.abs(slice - 10) <= 2
+
+
+def test_compute_midband_power():
+    """Test the compute_midband_power function with torch tensors."""
+    # Test parameters
+    ps = 6.5 / 100
+    lambda_ill = 0.532
+    NA_det = 1.4
+    midband_fractions = (0.125, 0.25)
+
+    # Create test data
+    np.random.seed(42)
+    test_2d_np = np.random.random((64, 64)).astype(np.float32)
+    test_2d_torch = torch.from_numpy(test_2d_np)
+
+    # Test the compute_midband_power function
+    result = focus.compute_midband_power(
+        test_2d_torch, NA_det, lambda_ill, ps, midband_fractions
+    )
+
+    # Check result properties
+    assert isinstance(result, torch.Tensor)
+    assert result.shape == torch.Size([])  # scalar tensor
+    assert result.item() > 0  # should be positive
+
+    # Test with different midband fractions
+    result2 = focus.compute_midband_power(
+        test_2d_torch, NA_det, lambda_ill, ps, (0.1, 0.3)
+    )
+    assert isinstance(result2, torch.Tensor)
+    assert result2.item() > 0
+
+    # Results should be different for different bands
+    assert abs(result.item() - result2.item()) > 1e-6
+
+
+def test_compute_midband_power_consistency():
+    """Test that compute_midband_power is consistent with focus_from_transverse_band."""
+    # Test parameters
+    ps = 6.5 / 100
+    lambda_ill = 0.532
+    NA_det = 1.4
+    midband_fractions = (0.125, 0.25)
+
+    # Create 3D test data
+    np.random.seed(42)
+    test_3d = np.random.random((3, 32, 32)).astype(np.float32)
+
+    # Test focus_from_transverse_band still works
+    focus_slice = focus.focus_from_transverse_band(
+        test_3d, NA_det, lambda_ill, ps, midband_fractions
+    )
+
+    assert isinstance(focus_slice, (int, np.integer))
+    assert 0 <= focus_slice < test_3d.shape[0]
+
+    # Manually compute midband power for each slice
+    manual_powers = []
+    for z in range(test_3d.shape[0]):
+        power = focus.compute_midband_power(
+            torch.from_numpy(test_3d[z]),
+            NA_det,
+            lambda_ill,
+            ps,
+            midband_fractions,
+        )
+        manual_powers.append(power.item())
+
+    expected_focus_slice = np.argmax(manual_powers)
+    assert focus_slice == expected_focus_slice
diff --git a/waveorder/focus.py b/waveorder/focus.py
index 38128f3c..38fcb521 100644
--- a/waveorder/focus.py
+++ b/waveorder/focus.py
@@ -3,11 +3,53 @@
 
 import matplotlib.pyplot as plt
 import numpy as np
+import torch
 from scipy.signal import peak_widths
 
 from waveorder import util
 
 
+def compute_midband_power(
+    yx_array: torch.Tensor,
+    NA_det: float,
+    lambda_ill: float,
+    pixel_size: float,
+    midband_fractions: tuple[float, float] = (0.125, 0.25),
+) -> torch.Tensor:
+    """Compute midband spatial frequency power by summing over a 2D midband donut.
+
+    Parameters
+    ----------
+    yx_array : torch.Tensor
+        2D tensor in (Y, X) order.
+    NA_det : float
+        Detection NA.
+    lambda_ill : float
+        Illumination wavelength.
+        Units are arbitrary, but must match [pixel_size].
+    pixel_size : float
+        Object-space pixel size = camera pixel size / magnification.
+        Units are arbitrary, but must match [lambda_ill].
+    midband_fractions : tuple[float, float], optional
+        The minimum and maximum fraction of the cutoff frequency that define the midband.
+        Default is (0.125, 0.25).
+
+    Returns
+    -------
+    torch.Tensor
+        Sum of absolute FFT values in the midband region.
+    """
+    _, _, fxx, fyy = util.gen_coordinate(yx_array.shape, pixel_size)
+    frr = torch.tensor(np.sqrt(fxx**2 + fyy**2))
+    xy_abs_fft = torch.abs(torch.fft.fftn(yx_array))
+    cutoff = 2 * NA_det / lambda_ill
+    mask = torch.logical_and(
+        frr > cutoff * midband_fractions[0],
+        frr < cutoff * midband_fractions[1],
+    )
+    return torch.sum(xy_abs_fft[mask])
+
+
 def focus_from_transverse_band(
     zyx_array,
     NA_det,
@@ -79,24 +121,20 @@ def focus_from_transverse_band(
         )
         return 0
 
-    # Calculate coordinates
-    _, Y, X = zyx_array.shape
-    _, _, fxx, fyy = util.gen_coordinate((Y, X), pixel_size)
-    frr = np.sqrt(fxx**2 + fyy**2)
-
-    # Calculate fft
-    xy_abs_fft = np.abs(np.fft.fftn(zyx_array, axes=(1, 2)))
-
-    # Calculate midband mask
-    cutoff = 2 * NA_det / lambda_ill
-    midband_mask = np.logical_and(
-        frr > cutoff * midband_fractions[0],
-        frr < cutoff * midband_fractions[1],
+    # Calculate midband power for each slice
+    midband_sum = np.array(
+        [
+            compute_midband_power(
+                torch.from_numpy(zyx_array[z]),
+                NA_det,
+                lambda_ill,
+                pixel_size,
+                midband_fractions,
+            ).numpy()
+            for z in range(zyx_array.shape[0])
+        ]
     )
 
-    # Find slice index with min/max power in midband
-    midband_sum = np.sum(xy_abs_fft[:, midband_mask], axis=1)
-
     if polynomial_fit_order is None:
         peak_index = minmaxfunc(midband_sum)
     else: