diff --git a/paddle/fluid/inference/api/demo_ci/utils.h b/paddle/fluid/inference/api/demo_ci/utils.h
index 5ac00fd294f433be08ad2f058529fc4b35e60e19..2ff5d88300b54853443926f6abe9a504050f170e 100644
--- a/paddle/fluid/inference/api/demo_ci/utils.h
+++ b/paddle/fluid/inference/api/demo_ci/utils.h
@@ -71,7 +71,8 @@ Record ProcessALine(const std::string& line) {
   return record;
 }
 
-void CheckOutput(const std::string& referfile, const PaddleTensor& output) {
+void CheckOutput(const std::string& referfile, const PaddleTensor& output,
+                 const float epsilon = 1e-5) {
   std::string line;
   std::ifstream file(referfile);
   std::getline(file, line);
@@ -93,7 +94,7 @@ void CheckOutput(const std::string& referfile, const PaddleTensor& output) {
       for (size_t i = 0; i < numel; ++i) {
         CHECK_LT(
             fabs(static_cast<float*>(output.data.data())[i] - refer.data[i]),
-            1e-5);
+            epsilon);
       }
       break;
     }
diff --git a/paddle/fluid/inference/api/demo_ci/vis_demo.cc b/paddle/fluid/inference/api/demo_ci/vis_demo.cc
index 0b3257da92cd3facc7f8f10950ffcd9fa3008bc8..5148c8039e692facc3e22948e5e5faf1da83a65b 100644
--- a/paddle/fluid/inference/api/demo_ci/vis_demo.cc
+++ b/paddle/fluid/inference/api/demo_ci/vis_demo.cc
@@ -71,7 +71,7 @@ void Main(bool use_gpu) {
   auto& tensor = output.front();
 
   // compare with reference result
-  CheckOutput(FLAGS_refer, tensor);
+  CheckOutput(FLAGS_refer, tensor, 1e-3);
 
   // the analysis_output has some diff with native_output,
   // TODO(luotao): add CheckOutput for analysis_output later.