Make the behavior of convert_image_dtype more explicit in the docs.

Change: 111949505

1 files changed, 5 insertions, 4 deletions

diff --git a/tensorflow/python/ops/image_ops.py b/tensorflow/python/ops/image_ops.py
index 90b9e93d35..ee09ada0e6 100644
--- a/tensorflow/python/ops/image_ops.py
+++ b/tensorflow/python/ops/image_ops.py
@@ -885,10 +885,11 @@ def convert_image_dtype(image, dtype, saturate=False, name=None):
 
   Note that converting from floating point inputs to integer types may lead to
   over/underflow problems. Set saturate to `True` to avoid such problem in
-  problematic conversions. Saturation will clip the output into the allowed
-  range before performing a potentially dangerous cast (i.e. when casting from
-  a floating point to an integer type, or when casting from an signed to an
-  unsigned type).
+  problematic conversions. If enabled, saturation will clip the output into the
+  allowed range before performing a potentially dangerous cast (and only before
+  performing such a cast, i.e., when casting from a floating point to an integer
+  type, and when casting from a signed to an unsigned type; `saturate` has no
+  effect on casts between floats, or on casts that increase the type's range).
 
   Args:
     image: An image.