diff --git a/comfy_extras/nodes_latent.py b/comfy_extras/nodes_latent.py
index 1f93e34d6..91b5e7bdb 100644
--- a/comfy_extras/nodes_latent.py
+++ b/comfy_extras/nodes_latent.py
@@ -120,14 +120,14 @@ class LatentInterpolate(io.ComfyNode):
 
         s2 = reshape_latent_to(s1.shape, s2)
 
-        m1 = torch.linalg.vector_norm(s1, dim=(1))
-        m2 = torch.linalg.vector_norm(s2, dim=(1))
+        m1 = torch.linalg.vector_norm(s1, dim=1, keepdim=True)
+        m2 = torch.linalg.vector_norm(s2, dim=1, keepdim=True)
 
         s1 = torch.nan_to_num(s1 / m1)
         s2 = torch.nan_to_num(s2 / m2)
 
         t = (s1 * ratio + s2 * (1.0 - ratio))
-        mt = torch.linalg.vector_norm(t, dim=(1))
+        mt = torch.linalg.vector_norm(t, dim=1, keepdim=True)
         st = torch.nan_to_num(t / mt)
 
         samples_out["samples"] = st * (m1 * ratio + m2 * (1.0 - ratio))
diff --git a/tests-unit/comfy_extras_test/nodes_latent_test.py b/tests-unit/comfy_extras_test/nodes_latent_test.py
new file mode 100644
index 000000000..001bfe55e
--- /dev/null
+++ b/tests-unit/comfy_extras_test/nodes_latent_test.py
@@ -0,0 +1,42 @@
+from unittest.mock import patch, MagicMock
+
+import torch
+
+mock_nodes = MagicMock()
+mock_nodes.MAX_RESOLUTION = 16384
+mock_server = MagicMock()
+
+# comfy.model_management initializes the torch device at import and requires CUDA
+# (or --cpu), which the unit-test environment does not provide; stub it for the import.
+patched_modules = {"nodes": mock_nodes, "server": mock_server, "comfy.model_management": MagicMock()}
+with patch.dict("sys.modules", patched_modules):
+    # imported first to satisfy the nodes_post_processing <-> nodes_latent circular import
+    import comfy_extras.nodes_post_processing  # noqa: F401
+    from comfy_extras.nodes_latent import LatentInterpolate
+
+
+class TestLatentInterpolate:
+    def test_batch_size_one(self):
+        s = {"samples": torch.randn(1, 4, 16, 16)}
+        out = LatentInterpolate.execute(s, s, 0.5)[0]
+        assert out["samples"].shape == (1, 4, 16, 16)
+
+    def test_batched_latent_does_not_crash(self):
+        s1 = {"samples": torch.randn(2, 4, 16, 16)}
+        s2 = {"samples": torch.randn(2, 4, 16, 16)}
+        out = LatentInterpolate.execute(s1, s2, 0.5)[0]
+        assert out["samples"].shape == (2, 4, 16, 16)
+
+    def test_each_batch_element_is_interpolated_independently(self):
+        # normalization must be per latent, not mixed across the batch. batch == channels
+        # (4) is the case that does not crash but silently corrupts without the fix.
+        a = torch.randn(1, 4, 8, 8)
+        b = torch.randn(1, 4, 8, 8)
+        filler = torch.randn(3, 4, 8, 8)
+        single = LatentInterpolate.execute({"samples": a}, {"samples": b}, 0.3)[0]["samples"]
+        batched = LatentInterpolate.execute(
+            {"samples": torch.cat([a, filler])},
+            {"samples": torch.cat([b, filler])},
+            0.3,
+        )[0]["samples"]
+        assert torch.allclose(single[0], batched[0], atol=1e-5)