diff --git a/onnxruntime/contrib_ops/cpu/image_scaler.h b/onnxruntime/contrib_ops/cpu/image_scaler.h
index 865bca51f1e85..12e47670a8f62 100644
--- a/onnxruntime/contrib_ops/cpu/image_scaler.h
+++ b/onnxruntime/contrib_ops/cpu/image_scaler.h
@@ -36,7 +36,11 @@ class ImageScaler final : public OpKernel {
     const int64_t H = dims[2];
     const int64_t W = dims[3];
 
-    if (!bias_.empty() && bias_.size() != static_cast<size_t>(C)) {
+    // The loop below reads bias_[nc % C] for every channel, so the bias must have exactly one entry
+    // per channel. An empty bias is not a valid "no bias" state here: the constructor already fails
+    // when the attribute is absent, and a present-but-empty attribute would otherwise index an
+    // empty vector.
+    if (bias_.size() != static_cast<size_t>(C)) {
       return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT, "Bias size (", bias_.size(), ") does not match the number of channels (", C, ")");
     }
 
diff --git a/onnxruntime/contrib_ops/cuda/tensor/image_scaler.cc b/onnxruntime/contrib_ops/cuda/tensor/image_scaler.cc
index befad5661c43f..8c0b60bedb6d5 100644
--- a/onnxruntime/contrib_ops/cuda/tensor/image_scaler.cc
+++ b/onnxruntime/contrib_ops/cuda/tensor/image_scaler.cc
@@ -47,7 +47,9 @@ Status ImageScaler<T>::ComputeInternal(OpKernelContext* context) const {
 
   const int64_t C = dims[1];  // dims are NCHW
 
-  if (!bias_.empty() && bias_.size() != static_cast<size_t>(C)) {
+  // The kernel indexes bias_data[c] for every channel, so the bias must have exactly one entry per
+  // channel. An empty bias would leave b_data_ pointing at a zero-sized allocation.
+  if (bias_.size() != static_cast<size_t>(C)) {
     return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT, "Bias size (", bias_.size(), ") does not match the number of channels (", C, ")");
   }
 
diff --git a/onnxruntime/test/contrib_ops/tensor_op_test.cc b/onnxruntime/test/contrib_ops/tensor_op_test.cc
index 2922e6517943e..d3d2b4a7d330a 100644
--- a/onnxruntime/test/contrib_ops/tensor_op_test.cc
+++ b/onnxruntime/test/contrib_ops/tensor_op_test.cc
@@ -98,6 +98,31 @@ TEST(ImageScalerContribOpTest, ImageScalerTest) {
   test.Run();
 }
 
+TEST(ImageScalerContribOpTest, ImageScalerEmptyBias) {
+  if (DefaultDmlExecutionProvider().get() != nullptr) {
+    GTEST_SKIP() << "Skipping because of the following error: AbiCustomRegistry.cpp(507): The parameter is incorrect.";
+  }
+
+  constexpr int64_t N = 1, C = 2, H = 2, W = 2;
+  std::vector<float> X = {
+      1.0f, 3.0f,
+      3.0f, 5.0f,
+
+      3.0f, 5.0f,
+      7.0f, 9.0f};
+
+  // A present-but-empty bias supplies no value for any channel and must be rejected rather than
+  // indexing an empty vector.
+  OpTester test("ImageScaler");
+  test.AddAttribute("scale", 2.0f);
+  test.AddAttribute("bias", std::vector<float>{});
+  test.AddInput<float>("input", {N, C, H, W}, X);
+  test.AddOutput<float>("output", {N, C, H, W}, std::vector<float>(N * C * H * W, 0.0f));
+  test.Run(OpTester::ExpectResult::kExpectFailure,
+           "Bias size (0) does not match the number of channels (2)",
+           {kTensorrtExecutionProvider});
+}
+
 void MeanVarianceNormalizationAcrossChannels(bool across_channels, bool normalize_variance) {
   constexpr int64_t N = 2, C = 2, H = 2, W = 3;
   constexpr int64_t one = 1;