diff --git a/tensorflow_datasets/core/features/image_feature_test.py b/tensorflow_datasets/core/features/image_feature_test.py
index 40dd092be87..accbec34175 100644
--- a/tensorflow_datasets/core/features/image_feature_test.py
+++ b/tensorflow_datasets/core/features/image_feature_test.py
@@ -67,11 +67,18 @@ def make_none_fail():
 
 
 def _unsupported_images_for_pil(
-    np_dtype: Type[np.generic], failing_lib: Optional[LibWithImportError]
+    np_dtype: Type[np.generic],
+    failing_lib: Optional[LibWithImportError],
+    channels: int,
 ) -> bool:
   # PIL is used when OpenCV is not installed, but PIL doesn't support 16-bit
   # images.
-  return np_dtype == np.uint16 and failing_lib == LibWithImportError.CV2
+  if np_dtype == np.uint16 and failing_lib == LibWithImportError.CV2:
+    return True
+  # PIL also doesn't support multi-channel 16-bit images for thumbnails.
+  if np_dtype == np.uint16 and channels > 1:
+    return True
+  return False
 
 
 class ImageFeatureTest(
@@ -91,7 +98,7 @@ class ImageFeatureTest(
   def test_images(self, make_lib_fail, dtypes, channels):
     dtype, np_dtype = dtypes
     with make_lib_fail() as failing_lib:
-      if _unsupported_images_for_pil(np_dtype, failing_lib):
+      if _unsupported_images_for_pil(np_dtype, failing_lib, channels):
         return
       img = randint(256, size=(128, 100, channels), dtype=np_dtype)
       img_other_shape = randint(256, size=(64, 200, channels), dtype=np_dtype)
@@ -198,7 +205,7 @@ def test_images(self, make_lib_fail, dtypes, channels):
   def test_images_with_invalid_shape(self, make_lib_fail, dtypes, channels):
     dtype, np_dtype = dtypes
     with make_lib_fail() as failing_lib:
-      if _unsupported_images_for_pil(np_dtype, failing_lib):
+      if _unsupported_images_for_pil(np_dtype, failing_lib, channels):
         return
       invalid_number_of_dimensions = testing.FeatureExpectationItem(
           value=randint(256, size=(128, 128), dtype=np_dtype),