diff --git a/tests/test_instance_norm.py b/tests/test_instance_norm.py
index d42d0ed3..297e8bbe 100644
--- a/tests/test_instance_norm.py
+++ b/tests/test_instance_norm.py
@@ -19,7 +19,10 @@
 from tests import util
 
 
-@pytest.mark.parametrize("num_batches, y_size, x_size, use_input_stats", [(2, 8, 1024, True), (3, 5, 2000, False)])
+@pytest.mark.parametrize(
+    "num_batches, y_size, x_size, use_input_stats",
+    [(2, 8, 1024, True), (2, 2, 2, True), (3, 5, 2000, False)],
+)
 def test_forward(num_batches, y_size, x_size, use_input_stats, device):
     input = torch.randn(num_batches, y_size, x_size, device=device)
 
diff --git a/trident/kernel/instance_norm.py b/trident/kernel/instance_norm.py
index 6c879f0e..29a5a388 100644
--- a/trident/kernel/instance_norm.py
+++ b/trident/kernel/instance_norm.py
@@ -143,6 +143,7 @@ def forward_running_mean_running_var(
         running_var_ptr: tl.tensor,
         num_batches: tl.int32,
         y_size: tl.int32,
+        x_size: tl.int32,
         momentum: tl.constexpr,
         batch_block_size: tl.constexpr,
     ):
@@ -188,7 +189,7 @@ def forward_running_mean_running_var(
         running_mean = tl.load(running_mean_block_ptr, boundary_check=(0,))
         running_mean = mean * momentum + running_mean * (1 - momentum)
         running_var = tl.load(running_var_block_ptr, boundary_check=(0,))
-        running_var = var * momentum + running_var * (1 - momentum)
+        running_var = var * (x_size / (x_size - 1)) * momentum + running_var * (1 - momentum)
         tl.store(running_mean_block_ptr, running_mean, boundary_check=(0,))
         tl.store(running_var_block_ptr, running_var, boundary_check=(0,))
 
diff --git a/trident/operation/instance_norm.py b/trident/operation/instance_norm.py
index b8235efb..5e6f361b 100644
--- a/trident/operation/instance_norm.py
+++ b/trident/operation/instance_norm.py
@@ -96,7 +96,15 @@ def grid(meta):
 
             util.push_trace("kernel.InstanceNorm.forward_running_mean_running_var")
             kernel.InstanceNorm.forward_running_mean_running_var[grid](
-                mean, var, running_mean, running_var, num_batches, y_size, momentum, triton.next_power_of_2(num_batches)
+                mean,
+                var,
+                running_mean,
+                running_var,
+                num_batches,
+                y_size,
+                x_size,
+                momentum,
+                triton.next_power_of_2(num_batches),
             )
             util.pop_trace()