diff --git a/tensorflow/compiler/tests/image_ops_test.py b/tensorflow/compiler/tests/image_ops_test.py
index 34e2c8c8986..28a1dd9ebdb 100644
--- a/tensorflow/compiler/tests/image_ops_test.py
+++ b/tensorflow/compiler/tests/image_ops_test.py
@@ -560,6 +560,8 @@ class ResizeBilinearTest(parameterized.TestCase, xla_test.XLATestCase):
       ("72x72To456x456", 72, 72, 456, 456),
       ("86x86To456x456", 86, 86, 456, 456),
       ("100x100To456x456", 100, 100, 456, 456),
+      ("64x64To224x224", 64, 64, 224, 224),
+      ("224x224To224x224", 224, 224, 224, 224),
       # This test is disabled because it is very slow. It is slow because
       # 383 is prime, 383 and 2047 are coprime, and 2048 is large.
       # ("Disabled_384x72To2048x384", 384, 72, 2048, 384),
diff --git a/tensorflow/compiler/tf2xla/kernels/image_resize_ops.cc b/tensorflow/compiler/tf2xla/kernels/image_resize_ops.cc
index 9ee17d1aa9c..aaa322eef87 100644
--- a/tensorflow/compiler/tf2xla/kernels/image_resize_ops.cc
+++ b/tensorflow/compiler/tf2xla/kernels/image_resize_ops.cc
@@ -312,8 +312,11 @@ xla::XlaOp ResizeUsingDilationAndConvolution(
   }
 
   // Split convolutions into independent dimensions if they would be a very
-  // large kernel.
-  if (dims.kernel_size[0] * dims.kernel_size[1] < kMax2DKernelSize) {
+  // large kernel or if one or more of the dimensions are already equal.
+  bool decompose_resize =
+      in_size[0] == out_size[0] || in_size[1] == out_size[1] ||
+      dims.kernel_size[0] * dims.kernel_size[1] >= kMax2DKernelSize;
+  if (!decompose_resize) {
     xla::XlaOp kernel = MakeGeneralResizeKernel(builder, type, dims.kernel_size,
                                                 channels, is_kernel_bilinear);
     output =