diff --git a/examples/infer_water/CMakeLists.txt b/examples/infer_water/CMakeLists.txt
index c82816c912..bfa15e4a16 100644
--- a/examples/infer_water/CMakeLists.txt
+++ b/examples/infer_water/CMakeLists.txt
@@ -4,6 +4,11 @@ project(infer_water)
 # find DeePMD-kit
 find_package(DeePMD REQUIRED)
 
+# helper that generates graph.pb from the bundled test model
+add_executable(convert_model convert_model.c)
+# link DeePMD-kit C API
+target_link_libraries(convert_model PRIVATE DeePMD::deepmd_c)
+
 # C++ example
 add_executable(infer_water_cc infer_water.cpp)
 # link DeePMD-kit C++ API
diff --git a/examples/infer_water/README.md b/examples/infer_water/README.md
index 290dd8cd1d..4008116251 100644
--- a/examples/infer_water/README.md
+++ b/examples/infer_water/README.md
@@ -8,3 +8,26 @@ Build the project using
 cmake -DCMAKE_PREFIX_PATH=$deepmd_root .
 make
 ```
+
+Building only compiles the executables; it does not create the model file.
+The inference programs load a frozen model `graph.pb` from the current
+directory, so generate it first by running the `convert_model` helper, which
+converts the bundled test model `../../source/tests/infer/deeppot.pbtxt` into
+`graph.pb`:
+
+```sh
+./convert_model
+```
+
+Run `convert_model` from this directory so the relative path to the bundled
+model resolves. It requires `$deepmd_root` to be built with the TensorFlow
+backend, because `graph.pb` is a TensorFlow frozen model.
+
+Then run any of the inference examples:
+
+```sh
+./infer_water_cc
+./infer_water_c
+./infer_water_hpp
+./infer_water_nlist
+```