diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index 0008f9890..676a2c7e7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -226,8 +226,8 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
           m_outputRows = numext::ceil((m_input_rows_eff - m_patch_rows_eff + 1.f) / static_cast<float>(m_row_strides));
           m_outputCols = numext::ceil((m_input_cols_eff - m_patch_cols_eff + 1.f) / static_cast<float>(m_col_strides));
           // Calculate the padding
-          m_rowPaddingTop = ((m_outputRows - 1) * m_row_strides + m_patch_rows_eff - m_input_rows_eff) / 2;
-          m_colPaddingLeft = ((m_outputCols - 1) * m_col_strides + m_patch_cols_eff - m_input_cols_eff) / 2;
+          m_rowPaddingTop = numext::maxi<Index>(0, ((m_outputRows - 1) * m_row_strides + m_patch_rows_eff - m_input_rows_eff) / 2);
+          m_colPaddingLeft = numext::maxi<Index>(0, ((m_outputCols - 1) * m_col_strides + m_patch_cols_eff - m_input_cols_eff) / 2);
           break;
         case PADDING_SAME:
           m_outputRows = numext::ceil(m_input_rows_eff / static_cast<float>(m_row_strides));