diff --git a/CHANGELOG.md b/CHANGELOG.md
index d7362c938cb..6695a09db43 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -42,6 +42,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Fixed multiclass recall macro avg. ignore index ([#2710](https://github.com/Lightning-AI/torchmetrics/pull/2710))
 
 
+- Fixed handling zero division error in binary IoU (Jaccard index) calculation ([#2726](https://github.com/Lightning-AI/torchmetrics/pull/2726))
+
+
 - Correct the padding related calculation errors in SSIM ([#2721](https://github.com/Lightning-AI/torchmetrics/pull/2721))
 
 
diff --git a/src/torchmetrics/functional/classification/jaccard.py b/src/torchmetrics/functional/classification/jaccard.py
index 1d240df68af..dfddd68255f 100644
--- a/src/torchmetrics/functional/classification/jaccard.py
+++ b/src/torchmetrics/functional/classification/jaccard.py
@@ -67,7 +67,7 @@ def _jaccard_index_reduce(
         raise ValueError(f"The `average` has to be one of {allowed_average}, got {average}.")
     confmat = confmat.float()
     if average == "binary":
-        return confmat[1, 1] / (confmat[0, 1] + confmat[1, 0] + confmat[1, 1])
+        return _safe_divide(confmat[1, 1], (confmat[0, 1] + confmat[1, 0] + confmat[1, 1]), zero_division=zero_division)
 
     ignore_index_cond = ignore_index is not None and 0 <= ignore_index < confmat.shape[0]
     multilabel = confmat.ndim == 3
diff --git a/tests/unittests/classification/test_jaccard.py b/tests/unittests/classification/test_jaccard.py
index 6901868eac9..e7afdb557a6 100644
--- a/tests/unittests/classification/test_jaccard.py
+++ b/tests/unittests/classification/test_jaccard.py
@@ -26,6 +26,7 @@
     MultilabelJaccardIndex,
 )
 from torchmetrics.functional.classification.jaccard import (
+    _jaccard_index_reduce,
     binary_jaccard_index,
     multiclass_jaccard_index,
     multilabel_jaccard_index,
@@ -403,6 +404,26 @@ def test_corner_case():
     assert torch.allclose(res, out)
 
 
+def test_jaccard_index_zero_division():
+    """Issue: https://github.com/Lightning-AI/torchmetrics/issues/2658."""
+    # Test case where all pixels are background (zeros)
+    confmat = torch.tensor([[4, 0], [0, 0]])
+
+    # Test with zero_division=0.0
+    result = _jaccard_index_reduce(confmat, average="binary", zero_division=0.0)
+    assert result == 0.0, f"Expected 0.0, but got {result}"
+
+    # Test with zero_division=1.0
+    result = _jaccard_index_reduce(confmat, average="binary", zero_division=1.0)
+    assert result == 1.0, f"Expected 1.0, but got {result}"
+
+    # Test case with some foreground pixels
+    confmat = torch.tensor([[2, 1], [1, 1]])
+    result = _jaccard_index_reduce(confmat, average="binary", zero_division=0.0)
+    expected = 1 / 3
+    assert torch.isclose(result, torch.tensor(expected)), f"Expected {expected}, but got {result}"
+
+
 @pytest.mark.parametrize(
     ("metric", "kwargs"),
     [