diff --git a/tensorflow/lite/delegates/gpu/common/object_reader.cc b/tensorflow/lite/delegates/gpu/common/object_reader.cc
index 55a0aea01a1..f299232a8a7 100644
--- a/tensorflow/lite/delegates/gpu/common/object_reader.cc
+++ b/tensorflow/lite/delegates/gpu/common/object_reader.cc
@@ -37,7 +37,7 @@ absl::Status ObjectReader::ReadNonConstantTensor(
   }
 
   if (tensor_to_value->find(tensor_idx) == tensor_to_value->end()) {
-    const TfLiteTensor& tflite_tensor = context->tensors[tensor_idx];
+    TfLiteTensor& tflite_tensor = context->tensors[tensor_idx];
     if (tflite::IsConstantTensor(&tflite_tensor)) {
       return absl::InvalidArgumentError(absl::StrCat(
           "ReadNonConstantTensor: value is a constant tensor: ", tensor_idx));
@@ -58,6 +58,7 @@ absl::Status ObjectReader::ReadNonConstantTensor(
                 &fp_tensor_index) != kTfLiteOk) {
           return absl::InternalError("Could not add new tensor to graph");
         }
+
         // Remember this tensor for later.
         (*quant_conversion_map)[fp_tensor_index] = tensor_idx;
         (*quant_conversion_map)[tensor_idx] = fp_tensor_index;
@@ -67,6 +68,9 @@ absl::Status ObjectReader::ReadNonConstantTensor(
             ConvertTfLiteTensorToTensorRef(*fp_tflite_tensor, &value->tensor));
         value->tensor.ref = fp_tensor_index;
         value->quant_params.emplace();
+        // tflite_tensor from the outer scope is invalidated due to calling
+        // CreateNewTensorWithDifferentType
+        tflite_tensor = context->tensors[tensor_idx];
         RETURN_IF_ERROR(
             PopulateQuantParams(tflite_tensor, &value->quant_params.value()));
         (*tensor_to_value)[fp_tensor_index] = value;
diff --git a/tensorflow/lite/delegates/utils.cc b/tensorflow/lite/delegates/utils.cc
index 873cadc180f..289586c5346 100644
--- a/tensorflow/lite/delegates/utils.cc
+++ b/tensorflow/lite/delegates/utils.cc
@@ -29,8 +29,8 @@ TfLiteStatus CreateNewTensorWithDifferentType(TfLiteContext* context,
                                               TfLiteType new_type,
                                               TfLiteTensor** new_tensor,
                                               int* new_tensor_index) {
-  const TfLiteTensor& original_tensor = context->tensors[original_tensor_index];
   TF_LITE_ENSURE_STATUS(context->AddTensors(context, 1, new_tensor_index));
+  const TfLiteTensor& original_tensor = context->tensors[original_tensor_index];
   *new_tensor = &context->tensors[*new_tensor_index];
   (*new_tensor)->type = new_type;
   (*new_tensor)->allocation_type = kTfLiteArenaRw;
diff --git a/tensorflow/lite/delegates/utils.h b/tensorflow/lite/delegates/utils.h
index 12684fcb84a..a9fb67316fc 100644
--- a/tensorflow/lite/delegates/utils.h
+++ b/tensorflow/lite/delegates/utils.h
@@ -33,7 +33,8 @@ namespace tflite {
 namespace delegates {
 
 // Creates a new Read/Write tensor having the same shape as the original, but
-// with a different type.
+// with a different type. Note that this might void existing references to
+// tensors.
 TfLiteStatus CreateNewTensorWithDifferentType(TfLiteContext* context,
                                               const int original_tensor_index,
                                               TfLiteType new_type,