diff --git a/.gitignore b/.gitignore
index 249cc9b..dd22728 100644
--- a/.gitignore
+++ b/.gitignore
@@ -4,12 +4,9 @@
 .pytest_cache
 .python-version
 build/
-betterproto/tests/*.bin
-betterproto/tests/*_pb2.py
-betterproto/tests/*.py
-!betterproto/tests/generate.py
-!betterproto/tests/test_*.py
+betterproto/tests/output_*
 **/__pycache__
 dist
 **/*.egg-info
 output
+.idea
\ No newline at end of file
diff --git a/README.md b/README.md
index e971ee2..53a7cdd 100644
--- a/README.md
+++ b/README.md
@@ -311,10 +311,26 @@ $ pip install -e .
 
 There are two types of tests:
 
-1. Manually-written tests for some behavior of the library
-2. Proto files and JSON inputs for automated tests
+1. Standard tests
+2. Custom tests
 
-For #2, you can add a new `*.proto` file into the `betterproto/tests` directory along with a sample `*.json` input and it will get automatically picked up.
+#### Standard tests
+
+Adding a standard test case is easy.
+
+- Create a new directory `betterproto/tests/inputs/<name>`
+  - add `<name>.proto`  with a message called `Test`
+  - add `<name>.json` with some test data
+
+It will be picked up automatically when you run the tests.
+
+- See also: [Standard Tests Development Guide](betterproto/tests/README.md) 
+
+#### Custom tests
+
+Custom tests are found in `tests/test_*.py` and are run with pytest.
+
+#### Running
 
 Here's how to run the tests.
 
@@ -322,7 +338,7 @@ Here's how to run the tests.
 # Generate assets from sample .proto files
 $ pipenv run generate
 
-# Run the tests
+# Run all tests
 $ pipenv run test
 ```
 
@@ -340,6 +356,9 @@ $ pipenv run test
 - [x] Refs to nested types
 - [x] Imports in proto files
 - [x] Well-known Google types
+  - [ ] Support as request input
+  - [ ] Support as response output
+    - [ ] Automatically wrap/unwrap responses
 - [x] OneOf support
   - [x] Basic support on the wire
   - [x] Check which was set from the group
diff --git a/betterproto/plugin.bat b/betterproto/plugin.bat
new file mode 100644
index 0000000..9b837d7
--- /dev/null
+++ b/betterproto/plugin.bat
@@ -0,0 +1,2 @@
+@SET plugin_dir=%~dp0
+@python %plugin_dir%/plugin.py %*
\ No newline at end of file
diff --git a/betterproto/tests/README.md b/betterproto/tests/README.md
new file mode 100644
index 0000000..ea15758
--- /dev/null
+++ b/betterproto/tests/README.md
@@ -0,0 +1,75 @@
+# Standard Tests Development Guide
+
+Standard test cases are found in [betterproto/tests/inputs](inputs), where each subdirectory represents a testcase, that is verified in isolation.
+
+```
+inputs/
+   bool/
+   double/
+   int32/
+   ...
+```
+
+## Test case directory structure
+
+Each testcase has a `<name>.proto` file with a message called `Test`, a matching `.json` file and optionally a custom test file called `test_*.py`.
+
+```bash
+bool/
+  bool.proto
+  bool.json
+  test_bool.py  # optional
+```
+
+### proto
+
+`<name>.proto` &mdash; *The protobuf message to test*
+
+```protobuf
+syntax = "proto3";
+
+message Test {
+    bool value = 1;
+}
+```
+
+You can add multiple `.proto` files to the test case, as long as one file matches the directory name. 
+
+### json
+
+`<name>.json` &mdash; *Test-data to validate the message with*
+
+```json
+{
+  "value": true
+}
+```
+
+### pytest
+
+`test_<name>.py` &mdash; *Custom test to validate specific aspects of the generated class*
+
+```python
+from betterproto.tests.output_betterproto.bool.bool import Test
+
+def test_value():
+    message = Test()
+    assert not message.value, "Boolean is False by default"
+```
+
+## Standard tests
+
+The following tests are automatically executed for all cases:
+
+- [x] Can the generated python code imported?
+- [x] Can the generated message class be instantiated?
+- [x] Is the generated code compatible with the Google's `grpc_tools.protoc` implementation?
+
+## Running the tests
+
+- `pipenv run generate`
+  This generates
+  - `betterproto/tests/output_betterproto` &mdash; *the plugin generated python classes*
+  - `betterproto/tests/output_reference` &mdash; *reference implementation classes*
+- `pipenv run test`
+
diff --git a/betterproto/tests/generate.py b/betterproto/tests/generate.py
index 987f2d9..fdcb220 100644
--- a/betterproto/tests/generate.py
+++ b/betterproto/tests/generate.py
@@ -1,84 +1,60 @@
 #!/usr/bin/env python
 import os
+import sys
+from typing import Set
+
+from betterproto.tests.util import get_directories, inputs_path, output_path_betterproto, output_path_reference, \
+    protoc_plugin, protoc_reference
 
 # Force pure-python implementation instead of C++, otherwise imports
 # break things because we can't properly reset the symbol database.
 os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
 
-import importlib
-import json
-import subprocess
-import sys
-from typing import Generator, Tuple
 
-from google.protobuf import symbol_database
-from google.protobuf.descriptor_pool import DescriptorPool
-from google.protobuf.json_format import MessageToJson, Parse
+def generate(whitelist: Set[str]):
+    path_whitelist = {os.path.realpath(e) for e in whitelist if os.path.exists(e)}
+    name_whitelist = {e for e in whitelist if not os.path.exists(e)}
+
+    test_case_names = set(get_directories(inputs_path))
+
+    for test_case_name in sorted(test_case_names):
+        test_case_path = os.path.realpath(os.path.join(inputs_path, test_case_name))
+
+        if whitelist and test_case_path not in path_whitelist and test_case_name not in name_whitelist:
+            continue
+
+        case_output_dir_reference = os.path.join(output_path_reference, test_case_name)
+        case_output_dir_betterproto = os.path.join(output_path_betterproto, test_case_name)
+
+        print(f'Generating output for {test_case_name}')
+        os.makedirs(case_output_dir_reference, exist_ok=True)
+        os.makedirs(case_output_dir_betterproto, exist_ok=True)
+
+        protoc_reference(test_case_path, case_output_dir_reference)
+        protoc_plugin(test_case_path, case_output_dir_betterproto)
 
 
-root = os.path.dirname(os.path.realpath(__file__))
+HELP = "\n".join([
+    'Usage: python generate.py',
+    '       python generate.py [DIRECTORIES or NAMES]',
+    'Generate python classes for standard tests.',
+    '',
+    'DIRECTORIES    One or more relative or absolute directories of test-cases to generate classes for.',
+    '               python generate.py inputs/bool inputs/double inputs/enum',
+    '',
+    'NAMES          One or more test-case names to generate classes for.',
+    '               python generate.py bool double enums'
+])
 
 
-def get_files(end: str) -> Generator[str, None, None]:
-    for r, dirs, files in os.walk(root):
-        for filename in [f for f in files if f.endswith(end)]:
-            yield os.path.join(r, filename)
+def main():
+    if set(sys.argv).intersection({'-h', '--help'}):
+        print(HELP)
+        return
+    whitelist = set(sys.argv[1:])
 
-
-def get_base(filename: str) -> str:
-    return os.path.splitext(os.path.basename(filename))[0]
-
-
-def ensure_ext(filename: str, ext: str) -> str:
-    if not filename.endswith(ext):
-        return filename + ext
-    return filename
+    generate(whitelist)
 
 
 if __name__ == "__main__":
-    os.chdir(root)
-
-    if len(sys.argv) > 1:
-        proto_files = [ensure_ext(f, ".proto") for f in sys.argv[1:]]
-        bases = {get_base(f) for f in proto_files}
-        json_files = [
-            f for f in get_files(".json") if get_base(f).split("-")[0] in bases
-        ]
-    else:
-        proto_files = get_files(".proto")
-        json_files = get_files(".json")
-
-    for filename in proto_files:
-        print(f"Generating code for {os.path.basename(filename)}")
-        subprocess.run(
-            f"protoc --python_out=. {os.path.basename(filename)}", shell=True
-        )
-        subprocess.run(
-            f"protoc --plugin=protoc-gen-custom=../plugin.py --custom_out=. {os.path.basename(filename)}",
-            shell=True,
-        )
-
-    for filename in json_files:
-        # Reset the internal symbol database so we can import the `Test` message
-        # multiple times. Ugh.
-        sym = symbol_database.Default()
-        sym.pool = DescriptorPool()
-
-        parts = get_base(filename).split("-")
-        out = filename.replace(".json", ".bin")
-        print(f"Using {parts[0]}_pb2 to generate {os.path.basename(out)}")
-
-        imported = importlib.import_module(f"{parts[0]}_pb2")
-        input_json = open(filename).read()
-        parsed = Parse(input_json, imported.Test())
-        serialized = parsed.SerializeToString()
-        preserve = "casing" not in filename
-        serialized_json = MessageToJson(parsed, preserving_proto_field_name=preserve)
-
-        s_loaded = json.loads(serialized_json)
-        in_loaded = json.loads(input_json)
-
-        if s_loaded != in_loaded:
-            raise AssertionError("Expected JSON to be equal:", s_loaded, in_loaded)
-
-        open(out, "wb").write(serialized)
+    main()
diff --git a/betterproto/tests/bool.json b/betterproto/tests/inputs/bool/bool.json
similarity index 100%
rename from betterproto/tests/bool.json
rename to betterproto/tests/inputs/bool/bool.json
diff --git a/betterproto/tests/bool.proto b/betterproto/tests/inputs/bool/bool.proto
similarity index 100%
rename from betterproto/tests/bool.proto
rename to betterproto/tests/inputs/bool/bool.proto
diff --git a/betterproto/tests/inputs/bool/test_bool.py b/betterproto/tests/inputs/bool/test_bool.py
new file mode 100644
index 0000000..0d4daa6
--- /dev/null
+++ b/betterproto/tests/inputs/bool/test_bool.py
@@ -0,0 +1,6 @@
+from betterproto.tests.output_betterproto.bool.bool import Test
+
+
+def test_value():
+    message = Test()
+    assert not message.value, "Boolean is False by default"
diff --git a/betterproto/tests/bytes.json b/betterproto/tests/inputs/bytes/bytes.json
similarity index 100%
rename from betterproto/tests/bytes.json
rename to betterproto/tests/inputs/bytes/bytes.json
diff --git a/betterproto/tests/bytes.proto b/betterproto/tests/inputs/bytes/bytes.proto
similarity index 100%
rename from betterproto/tests/bytes.proto
rename to betterproto/tests/inputs/bytes/bytes.proto
diff --git a/betterproto/tests/casing.json b/betterproto/tests/inputs/casing/casing.json
similarity index 100%
rename from betterproto/tests/casing.json
rename to betterproto/tests/inputs/casing/casing.json
diff --git a/betterproto/tests/casing.proto b/betterproto/tests/inputs/casing/casing.proto
similarity index 64%
rename from betterproto/tests/casing.proto
rename to betterproto/tests/inputs/casing/casing.proto
index 4ab37ae..ad0c427 100644
--- a/betterproto/tests/casing.proto
+++ b/betterproto/tests/inputs/casing/casing.proto
@@ -9,4 +9,9 @@ enum my_enum {
 message Test {
   int32 camelCase = 1;
   my_enum snake_case = 2;
+  snake_case_message snake_case_message = 3;
 }
+
+message snake_case_message {
+
+}
\ No newline at end of file
diff --git a/betterproto/tests/inputs/casing/test_casing.py b/betterproto/tests/inputs/casing/test_casing.py
new file mode 100644
index 0000000..3d1ac3d
--- /dev/null
+++ b/betterproto/tests/inputs/casing/test_casing.py
@@ -0,0 +1,16 @@
+import betterproto.tests.output_betterproto.casing.casing as casing
+from betterproto.tests.output_betterproto.casing.casing import Test
+
+
+def test_message_attributes():
+    message = Test()
+    assert hasattr(message, 'snake_case_message'), 'snake_case field name is same in python'
+    assert hasattr(message, 'camel_case'), 'CamelCase field is snake_case in python'
+
+
+def test_message_casing():
+    assert hasattr(casing, 'SnakeCaseMessage'), 'snake_case Message name is converted to CamelCase in python'
+
+
+def test_enum_casing():
+    assert hasattr(casing, 'MyEnum'), 'snake_case Enum name is converted to CamelCase in python'
diff --git a/betterproto/tests/double-negative.json b/betterproto/tests/inputs/double/double-negative.json
similarity index 100%
rename from betterproto/tests/double-negative.json
rename to betterproto/tests/inputs/double/double-negative.json
diff --git a/betterproto/tests/double.json b/betterproto/tests/inputs/double/double.json
similarity index 100%
rename from betterproto/tests/double.json
rename to betterproto/tests/inputs/double/double.json
diff --git a/betterproto/tests/double.proto b/betterproto/tests/inputs/double/double.proto
similarity index 100%
rename from betterproto/tests/double.proto
rename to betterproto/tests/inputs/double/double.proto
diff --git a/betterproto/tests/enums.json b/betterproto/tests/inputs/enums/enums.json
similarity index 100%
rename from betterproto/tests/enums.json
rename to betterproto/tests/inputs/enums/enums.json
diff --git a/betterproto/tests/enums.proto b/betterproto/tests/inputs/enums/enums.proto
similarity index 100%
rename from betterproto/tests/enums.proto
rename to betterproto/tests/inputs/enums/enums.proto
diff --git a/betterproto/tests/googletypes-missing.json b/betterproto/tests/inputs/googletypes/googletypes-missing.json
similarity index 100%
rename from betterproto/tests/googletypes-missing.json
rename to betterproto/tests/inputs/googletypes/googletypes-missing.json
diff --git a/betterproto/tests/googletypes.json b/betterproto/tests/inputs/googletypes/googletypes.json
similarity index 100%
rename from betterproto/tests/googletypes.json
rename to betterproto/tests/inputs/googletypes/googletypes.json
diff --git a/betterproto/tests/googletypes.proto b/betterproto/tests/inputs/googletypes/googletypes.proto
similarity index 100%
rename from betterproto/tests/googletypes.proto
rename to betterproto/tests/inputs/googletypes/googletypes.proto
diff --git a/betterproto/tests/inputs/googletypes_response/googletypes_response.proto b/betterproto/tests/inputs/googletypes_response/googletypes_response.proto
new file mode 100644
index 0000000..4bdca68
--- /dev/null
+++ b/betterproto/tests/inputs/googletypes_response/googletypes_response.proto
@@ -0,0 +1,18 @@
+syntax = "proto3";
+
+import "google/protobuf/wrappers.proto";
+
+service Test {
+    rpc GetInt32 (Input) returns (google.protobuf.Int32Value);
+    rpc GetAnotherInt32 (Input) returns (google.protobuf.Int32Value);
+    rpc GetInt64 (Input) returns (google.protobuf.Int64Value);
+    rpc GetOutput (Input) returns (Output);
+}
+
+message Input {
+
+}
+
+message Output {
+    google.protobuf.Int64Value int64 = 1;
+}
\ No newline at end of file
diff --git a/betterproto/tests/inputs/googletypes_response/test_googletypes_response.py b/betterproto/tests/inputs/googletypes_response/test_googletypes_response.py
new file mode 100644
index 0000000..9e8e454
--- /dev/null
+++ b/betterproto/tests/inputs/googletypes_response/test_googletypes_response.py
@@ -0,0 +1,18 @@
+from typing import Optional
+
+import pytest
+
+from betterproto.tests.output_betterproto.googletypes_response.googletypes_response import TestStub
+
+
+class TestStubChild(TestStub):
+    async def _unary_unary(self, route, request, response_type, **kwargs):
+        self.response_type = response_type
+
+
+@pytest.mark.asyncio
+async def test():
+    pytest.skip('todo')
+    stub = TestStubChild(None)
+    await stub.get_int64()
+    assert stub.response_type != Optional[int]
diff --git a/betterproto/tests/inputs/int32/int32.json b/betterproto/tests/inputs/int32/int32.json
new file mode 100644
index 0000000..34d4111
--- /dev/null
+++ b/betterproto/tests/inputs/int32/int32.json
@@ -0,0 +1,4 @@
+{
+  "positive": 150,
+  "negative": -150
+}
diff --git a/betterproto/tests/int32.proto b/betterproto/tests/inputs/int32/int32.proto
similarity index 72%
rename from betterproto/tests/int32.proto
rename to betterproto/tests/inputs/int32/int32.proto
index 6b46857..cae0dc7 100644
--- a/betterproto/tests/int32.proto
+++ b/betterproto/tests/inputs/int32/int32.proto
@@ -3,5 +3,6 @@ syntax = "proto3";
 // Some documentation about the Test message.
 message Test {
     // Some documentation about the count.
-    int32 count = 1;
+    int32 positive = 1;
+    int32 negative = 2;
 }
diff --git a/betterproto/tests/keywords.json b/betterproto/tests/inputs/keywords/keywords.json
similarity index 100%
rename from betterproto/tests/keywords.json
rename to betterproto/tests/inputs/keywords/keywords.json
diff --git a/betterproto/tests/keywords.proto b/betterproto/tests/inputs/keywords/keywords.proto
similarity index 100%
rename from betterproto/tests/keywords.proto
rename to betterproto/tests/inputs/keywords/keywords.proto
diff --git a/betterproto/tests/map.json b/betterproto/tests/inputs/map/map.json
similarity index 100%
rename from betterproto/tests/map.json
rename to betterproto/tests/inputs/map/map.json
diff --git a/betterproto/tests/map.proto b/betterproto/tests/inputs/map/map.proto
similarity index 100%
rename from betterproto/tests/map.proto
rename to betterproto/tests/inputs/map/map.proto
diff --git a/betterproto/tests/mapmessage.json b/betterproto/tests/inputs/mapmessage/mapmessage.json
similarity index 100%
rename from betterproto/tests/mapmessage.json
rename to betterproto/tests/inputs/mapmessage/mapmessage.json
diff --git a/betterproto/tests/mapmessage.proto b/betterproto/tests/inputs/mapmessage/mapmessage.proto
similarity index 100%
rename from betterproto/tests/mapmessage.proto
rename to betterproto/tests/inputs/mapmessage/mapmessage.proto
diff --git a/betterproto/tests/nested.json b/betterproto/tests/inputs/nested/nested.json
similarity index 100%
rename from betterproto/tests/nested.json
rename to betterproto/tests/inputs/nested/nested.json
diff --git a/betterproto/tests/nested.proto b/betterproto/tests/inputs/nested/nested.proto
similarity index 100%
rename from betterproto/tests/nested.proto
rename to betterproto/tests/inputs/nested/nested.proto
diff --git a/betterproto/tests/nestedtwice.json b/betterproto/tests/inputs/nestedtwice/nestedtwice.json
similarity index 100%
rename from betterproto/tests/nestedtwice.json
rename to betterproto/tests/inputs/nestedtwice/nestedtwice.json
diff --git a/betterproto/tests/nestedtwice.proto b/betterproto/tests/inputs/nestedtwice/nestedtwice.proto
similarity index 100%
rename from betterproto/tests/nestedtwice.proto
rename to betterproto/tests/inputs/nestedtwice/nestedtwice.proto
diff --git a/betterproto/tests/oneof-name.json b/betterproto/tests/inputs/oneof/oneof-name.json
similarity index 100%
rename from betterproto/tests/oneof-name.json
rename to betterproto/tests/inputs/oneof/oneof-name.json
diff --git a/betterproto/tests/oneof.json b/betterproto/tests/inputs/oneof/oneof.json
similarity index 100%
rename from betterproto/tests/oneof.json
rename to betterproto/tests/inputs/oneof/oneof.json
diff --git a/betterproto/tests/oneof.proto b/betterproto/tests/inputs/oneof/oneof.proto
similarity index 100%
rename from betterproto/tests/oneof.proto
rename to betterproto/tests/inputs/oneof/oneof.proto
diff --git a/betterproto/tests/ref.json b/betterproto/tests/inputs/ref/ref.json
similarity index 100%
rename from betterproto/tests/ref.json
rename to betterproto/tests/inputs/ref/ref.json
diff --git a/betterproto/tests/ref.proto b/betterproto/tests/inputs/ref/ref.proto
similarity index 100%
rename from betterproto/tests/ref.proto
rename to betterproto/tests/inputs/ref/ref.proto
diff --git a/betterproto/tests/repeatedmessage.proto b/betterproto/tests/inputs/ref/repeatedmessage.proto
similarity index 100%
rename from betterproto/tests/repeatedmessage.proto
rename to betterproto/tests/inputs/ref/repeatedmessage.proto
diff --git a/betterproto/tests/repeated.json b/betterproto/tests/inputs/repeated/repeated.json
similarity index 100%
rename from betterproto/tests/repeated.json
rename to betterproto/tests/inputs/repeated/repeated.json
diff --git a/betterproto/tests/repeated.proto b/betterproto/tests/inputs/repeated/repeated.proto
similarity index 100%
rename from betterproto/tests/repeated.proto
rename to betterproto/tests/inputs/repeated/repeated.proto
diff --git a/betterproto/tests/repeatedmessage.json b/betterproto/tests/inputs/repeatedmessage/repeatedmessage.json
similarity index 100%
rename from betterproto/tests/repeatedmessage.json
rename to betterproto/tests/inputs/repeatedmessage/repeatedmessage.json
diff --git a/betterproto/tests/inputs/repeatedmessage/repeatedmessage.proto b/betterproto/tests/inputs/repeatedmessage/repeatedmessage.proto
new file mode 100644
index 0000000..0ffacaf
--- /dev/null
+++ b/betterproto/tests/inputs/repeatedmessage/repeatedmessage.proto
@@ -0,0 +1,11 @@
+syntax = "proto3";
+
+package repeatedmessage;
+
+message Test {
+  repeated Sub greetings = 1;
+}
+
+message Sub {
+  string greeting = 1;
+}
\ No newline at end of file
diff --git a/betterproto/tests/repeatedpacked.json b/betterproto/tests/inputs/repeatedpacked/repeatedpacked.json
similarity index 100%
rename from betterproto/tests/repeatedpacked.json
rename to betterproto/tests/inputs/repeatedpacked/repeatedpacked.json
diff --git a/betterproto/tests/repeatedpacked.proto b/betterproto/tests/inputs/repeatedpacked/repeatedpacked.proto
similarity index 100%
rename from betterproto/tests/repeatedpacked.proto
rename to betterproto/tests/inputs/repeatedpacked/repeatedpacked.proto
diff --git a/betterproto/tests/service.proto b/betterproto/tests/inputs/service/service.proto
similarity index 100%
rename from betterproto/tests/service.proto
rename to betterproto/tests/inputs/service/service.proto
diff --git a/betterproto/tests/inputs/signed/signed.json b/betterproto/tests/inputs/signed/signed.json
new file mode 100644
index 0000000..b171e15
--- /dev/null
+++ b/betterproto/tests/inputs/signed/signed.json
@@ -0,0 +1,6 @@
+{
+  "signed32": 150,
+  "negative32": -150,
+  "string64": "150",
+  "negative64": "-150"
+}
diff --git a/betterproto/tests/inputs/signed/signed.proto b/betterproto/tests/inputs/signed/signed.proto
new file mode 100644
index 0000000..23fc9ee
--- /dev/null
+++ b/betterproto/tests/inputs/signed/signed.proto
@@ -0,0 +1,9 @@
+syntax = "proto3";
+
+message Test {
+    // todo: rename fields after fixing bug where 'signed_32_positive' will map to 'signed_32Positive' as output json
+    sint32 signed32 = 1;    //  signed_32_positive
+    sint32 negative32 = 2;  //  signed_32_negative
+    sint64 string64 = 3;    //  signed_64_positive
+    sint64 negative64 = 4;  //  signed_64_negative
+}
diff --git a/betterproto/tests/int32-negative.json b/betterproto/tests/int32-negative.json
deleted file mode 100644
index 0d2bb48..0000000
--- a/betterproto/tests/int32-negative.json
+++ /dev/null
@@ -1,3 +0,0 @@
-{
-  "count": -150
-}
diff --git a/betterproto/tests/int32.json b/betterproto/tests/int32.json
deleted file mode 100644
index 9514828..0000000
--- a/betterproto/tests/int32.json
+++ /dev/null
@@ -1,3 +0,0 @@
-{
-  "count": 150
-}
diff --git a/betterproto/tests/signed-negative.json b/betterproto/tests/signed-negative.json
deleted file mode 100644
index 2f6525a..0000000
--- a/betterproto/tests/signed-negative.json
+++ /dev/null
@@ -1,4 +0,0 @@
-{
-  "signed_32": -150,
-  "signed_64": "-150"
-}
diff --git a/betterproto/tests/signed.json b/betterproto/tests/signed.json
deleted file mode 100644
index 6049d88..0000000
--- a/betterproto/tests/signed.json
+++ /dev/null
@@ -1,4 +0,0 @@
-{
-  "signed_32": 150,
-  "signed_64": "150"
-}
diff --git a/betterproto/tests/signed.proto b/betterproto/tests/signed.proto
deleted file mode 100644
index 49b2bfd..0000000
--- a/betterproto/tests/signed.proto
+++ /dev/null
@@ -1,6 +0,0 @@
-syntax = "proto3";
-
-message Test {
-    sint32 signed_32 = 1;
-    sint64 signed_64 = 2;
-}
diff --git a/betterproto/tests/test_inputs.py b/betterproto/tests/test_inputs.py
index e8bc66c..d819b2e 100644
--- a/betterproto/tests/test_inputs.py
+++ b/betterproto/tests/test_inputs.py
@@ -1,32 +1,99 @@
 import importlib
 import json
-
+import os
+import sys
 import pytest
+import betterproto
+from betterproto.tests.util import get_directories, inputs_path
 
-from .generate import get_base, get_files
+# Force pure-python implementation instead of C++, otherwise imports
+# break things because we can't properly reset the symbol database.
+os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
 
-inputs = get_files(".bin")
+from google.protobuf import symbol_database
+from google.protobuf.descriptor_pool import DescriptorPool
+from google.protobuf.json_format import Parse
 
 
-@pytest.mark.parametrize("filename", inputs)
-def test_sample(filename: str) -> None:
-    module = get_base(filename).split("-")[0]
-    imported = importlib.import_module(f"betterproto.tests.{module}")
-    data_binary = open(filename, "rb").read()
-    data_dict = json.loads(open(filename.replace(".bin", ".json")).read())
-    t1 = imported.Test().parse(data_binary)
-    t2 = imported.Test().from_dict(data_dict)
-    print(t1)
-    print(t2)
+excluded_test_cases = {'googletypes_response', 'service'}
+test_case_names = {*get_directories(inputs_path)} - excluded_test_cases
 
-    # Equality should automagically work for dataclasses!
-    assert t1 == t2
+plugin_output_package = 'betterproto.tests.output_betterproto'
+reference_output_package = 'betterproto.tests.output_reference'
 
-    # Generally this can't be relied on, but here we are aiming to match the
-    # existing Python implementation and aren't doing anything tricky.
-    # https://developers.google.com/protocol-buffers/docs/encoding#implications
-    assert bytes(t1) == data_binary
-    assert bytes(t2) == data_binary
 
-    assert t1.to_dict() == data_dict
-    assert t2.to_dict() == data_dict
+@pytest.mark.parametrize("test_case_name", test_case_names)
+def test_message_can_be_imported(test_case_name: str) -> None:
+    importlib.import_module(f"{plugin_output_package}.{test_case_name}.{test_case_name}")
+
+
+@pytest.mark.parametrize("test_case_name", test_case_names)
+def test_message_can_instantiated(test_case_name: str) -> None:
+    plugin_module = importlib.import_module(f"{plugin_output_package}.{test_case_name}.{test_case_name}")
+    plugin_module.Test()
+
+
+@pytest.mark.parametrize("test_case_name", test_case_names)
+def test_message_equality(test_case_name: str) -> None:
+    plugin_module = importlib.import_module(f"{plugin_output_package}.{test_case_name}.{test_case_name}")
+    message1 = plugin_module.Test()
+    message2 = plugin_module.Test()
+    assert message1 == message2
+
+
+@pytest.mark.parametrize("test_case_name", test_case_names)
+def test_message_json(test_case_name: str) -> None:
+    plugin_module = importlib.import_module(f"{plugin_output_package}.{test_case_name}.{test_case_name}")
+    message: betterproto.Message = plugin_module.Test()
+    reference_json_data = get_test_case_json_data(test_case_name)
+
+    message.from_json(reference_json_data)
+    message_json = message.to_json(0)
+
+    assert json.loads(reference_json_data) == json.loads(message_json)
+
+
+@pytest.mark.parametrize("test_case_name", test_case_names)
+def test_binary_compatibility(test_case_name: str) -> None:
+    # Reset the internal symbol database so we can import the `Test` message
+    # multiple times. Ugh.
+    sym = symbol_database.Default()
+    sym.pool = DescriptorPool()
+
+    reference_module_root = os.path.join(*reference_output_package.split('.'), test_case_name)
+
+    sys.path.append(reference_module_root)
+
+    # import reference message
+    reference_module = importlib.import_module(f"{reference_output_package}.{test_case_name}.{test_case_name}_pb2")
+    plugin_module = importlib.import_module(f"{plugin_output_package}.{test_case_name}.{test_case_name}")
+
+    test_data = get_test_case_json_data(test_case_name)
+
+    reference_instance = Parse(test_data, reference_module.Test())
+    reference_binary_output = reference_instance.SerializeToString()
+
+    plugin_instance_from_json: betterproto.Message = plugin_module.Test().from_json(test_data)
+    plugin_instance_from_binary = plugin_module.Test.FromString(reference_binary_output)
+
+    # # Generally this can't be relied on, but here we are aiming to match the
+    # # existing Python implementation and aren't doing anything tricky.
+    # # https://developers.google.com/protocol-buffers/docs/encoding#implications
+    assert plugin_instance_from_json == plugin_instance_from_binary
+    assert plugin_instance_from_json.to_dict() == plugin_instance_from_binary.to_dict()
+
+    sys.path.remove(reference_module_root)
+
+
+'''
+helper methods
+'''
+
+
+def get_test_case_json_data(test_case_name):
+    test_data_path = os.path.join(inputs_path, test_case_name, f'{test_case_name}.json')
+    if not os.path.exists(test_data_path):
+        return None
+
+    with open(test_data_path) as fh:
+        return fh.read()
diff --git a/betterproto/tests/test_service_stub.py b/betterproto/tests/test_service_stub.py
index a5ba200..86377e2 100644
--- a/betterproto/tests/test_service_stub.py
+++ b/betterproto/tests/test_service_stub.py
@@ -3,8 +3,8 @@ import grpclib
 from grpclib.testing import ChannelFor
 import pytest
 from typing import Dict
-from .service import DoThingResponse, DoThingRequest, ExampleServiceStub
 
+from betterproto.tests.output_betterproto.service.service import DoThingResponse, DoThingRequest, ExampleServiceStub
 
 class ExampleService:
     def __init__(self, test_hook=None):
diff --git a/betterproto/tests/util.py b/betterproto/tests/util.py
new file mode 100644
index 0000000..b627a23
--- /dev/null
+++ b/betterproto/tests/util.py
@@ -0,0 +1,48 @@
+import os
+import subprocess
+from typing import Generator
+
+os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+
+root_path = os.path.dirname(os.path.realpath(__file__))
+inputs_path = os.path.join(root_path, 'inputs')
+output_path_reference = os.path.join(root_path, 'output_reference')
+output_path_betterproto = os.path.join(root_path, 'output_betterproto')
+
+if os.name == 'nt':
+    plugin_path = os.path.join(root_path, '..', 'plugin.bat')
+else:
+    plugin_path = os.path.join(root_path, '..', 'plugin.py')
+
+
+def get_files(path, end: str) -> Generator[str, None, None]:
+    for r, dirs, files in os.walk(path):
+        for filename in [f for f in files if f.endswith(end)]:
+            yield os.path.join(r, filename)
+
+
+def get_directories(path):
+    for root, directories, files in os.walk(path):
+        for directory in directories:
+            yield directory
+
+
+def relative(file: str, path: str):
+    return os.path.join(os.path.dirname(file), path)
+
+
+def read_relative(file: str, path: str):
+    with open(relative(file, path)) as fh:
+        return fh.read()
+
+
+def protoc_plugin(path: str, output_dir: str):
+    subprocess.run(
+        f"protoc --plugin=protoc-gen-custom={plugin_path} --custom_out={output_dir} --proto_path={path} {path}/*.proto",
+        shell=True,
+    )
+
+
+def protoc_reference(path: str, output_dir: str):
+    subprocess.run(f"protoc --python_out={output_dir} --proto_path={path} {path}/*.proto", shell=True)
+
diff --git a/pytest.ini b/pytest.ini
new file mode 100644
index 0000000..bec2b96
--- /dev/null
+++ b/pytest.ini
@@ -0,0 +1,5 @@
+[pytest]
+python_files = test_*.py
+python_classes =
+norecursedirs = **/output_*
+addopts = -p no:warnings
\ No newline at end of file