diff --git a/nodes/generator.py b/nodes/generator.py
index 3704373..32a02a9 100644
--- a/nodes/generator.py
+++ b/nodes/generator.py
@@ -1,2 +1,62 @@
+import tempfile
+import os
+import torch
+import torchaudio
+
+
 class OmniVoiceGenerate:
-    pass
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "model": ("OMNIVOICE_MODEL",),
+                "text": ("STRING", {"multiline": True, "default": ""}),
+                "mode": (
+                    ["voice_cloning", "voice_design", "auto_voice"],
+                    {"default": "voice_cloning"},
+                ),
+            },
+            "optional": {
+                "ref_audio": ("AUDIO",),
+                "ref_text": ("STRING", {"default": ""}),
+                "instruct": ("STRING", {"default": ""}),
+                "speed": ("FLOAT", {"default": 1.0, "min": 0.1, "max": 3.0, "step": 0.1}),
+                "num_step": ("INT", {"default": 32, "min": 1, "max": 100}),
+            },
+        }
+
+    RETURN_TYPES = ("AUDIO",)
+    RETURN_NAMES = ("audio",)
+    FUNCTION = "generate"
+    CATEGORY = "OmniVoice"
+
+    def generate(self, model, text, mode, ref_audio=None, ref_text="", instruct="", speed=1.0, num_step=32):
+        kwargs = {"text": text, "speed": speed, "num_step": num_step}
+
+        if mode == "voice_cloning" and ref_audio is not None:
+            tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+            tmp_path = tmp.name
+            tmp.close()
+            try:
+                waveform = ref_audio["waveform"].squeeze(0)  # (channels, samples)
+                torchaudio.save(tmp_path, waveform, ref_audio["sample_rate"])
+                kwargs["ref_audio"] = tmp_path
+                if ref_text:
+                    kwargs["ref_text"] = ref_text
+                audio_tensors = model.generate(**kwargs)
+            finally:
+                os.unlink(tmp_path)
+
+        elif mode == "voice_design" and instruct:
+            kwargs["instruct"] = instruct
+            audio_tensors = model.generate(**kwargs)
+
+        else:  # auto_voice or fallback
+            audio_tensors = model.generate(**kwargs)
+
+        # Concatenate chunks: each tensor is (1, T) → concat along T → (1, T_total)
+        combined = torch.cat(audio_tensors, dim=1)  # (1, T_total)
+        # ComfyUI AUDIO format: (batch, channels, samples)
+        waveform = combined.unsqueeze(0)  # (1, 1, T_total)
+
+        return ({"waveform": waveform, "sample_rate": 24000},)
diff --git a/tests/test_generator.py b/tests/test_generator.py
new file mode 100644
index 0000000..eda3a12
--- /dev/null
+++ b/tests/test_generator.py
@@ -0,0 +1,106 @@
+# tests/test_generator.py
+from unittest.mock import patch, MagicMock, call
+import torch
+import pytest
+from nodes.generator import OmniVoiceGenerate
+
+
+def make_mock_model(return_tensors=None):
+    mock = MagicMock()
+    if return_tensors is None:
+        return_tensors = [torch.zeros(1, 24000)]  # 1 second of silence
+    mock.generate.return_value = return_tensors
+    return mock
+
+
+def test_input_types_structure():
+    inputs = OmniVoiceGenerate.INPUT_TYPES()
+    required = inputs["required"]
+    assert "model" in required
+    assert "text" in required
+    assert "mode" in required
+    optional = inputs.get("optional", {})
+    assert "ref_audio" in optional
+    assert "ref_text" in optional
+    assert "instruct" in optional
+    assert "speed" in optional
+    assert "num_step" in optional
+
+
+def test_return_type():
+    assert OmniVoiceGenerate.RETURN_TYPES == ("AUDIO",)
+
+
+def test_generate_auto_voice():
+    node = OmniVoiceGenerate()
+    mock_model = make_mock_model()
+    result = node.generate(
+        model=mock_model,
+        text="Hello world",
+        mode="auto_voice",
+        speed=1.0,
+        num_step=32,
+    )
+    audio = result[0]
+    assert "waveform" in audio
+    assert "sample_rate" in audio
+    assert audio["sample_rate"] == 24000
+    mock_model.generate.assert_called_once_with(
+        text="Hello world", speed=1.0, num_step=32
+    )
+
+
+def test_generate_voice_design():
+    node = OmniVoiceGenerate()
+    mock_model = make_mock_model()
+    result = node.generate(
+        model=mock_model,
+        text="Hello world",
+        mode="voice_design",
+        instruct="female, low pitch",
+        speed=1.0,
+        num_step=32,
+    )
+    audio = result[0]
+    assert audio["sample_rate"] == 24000
+    mock_model.generate.assert_called_once_with(
+        text="Hello world", instruct="female, low pitch", speed=1.0, num_step=32
+    )
+
+
+def test_generate_voice_cloning():
+    node = OmniVoiceGenerate()
+    mock_model = make_mock_model()
+    # Simulate ComfyUI AUDIO input: waveform shape (batch, channels, samples)
+    ref_waveform = torch.zeros(1, 1, 24000)
+    ref_audio_input = {"waveform": ref_waveform, "sample_rate": 24000}
+
+    with patch("nodes.generator.torchaudio.save") as mock_save:
+        result = node.generate(
+            model=mock_model,
+            text="Hello world",
+            mode="voice_cloning",
+            ref_audio=ref_audio_input,
+            ref_text="reference text",
+            speed=1.0,
+            num_step=32,
+        )
+
+    assert mock_save.called
+    call_kwargs = mock_model.generate.call_args[1]
+    assert call_kwargs["ref_text"] == "reference text"
+    assert "ref_audio" in call_kwargs
+
+
+def test_output_waveform_shape():
+    node = OmniVoiceGenerate()
+    # Simulate two chunks returned by OmniVoice
+    chunk1 = torch.zeros(1, 24000)
+    chunk2 = torch.zeros(1, 12000)
+    mock_model = make_mock_model(return_tensors=[chunk1, chunk2])
+    result = node.generate(
+        model=mock_model, text="Long text", mode="auto_voice", speed=1.0, num_step=32
+    )
+    waveform = result[0]["waveform"]
+    # Shape must be (batch=1, channels=1, samples=36000)
+    assert waveform.shape == (1, 1, 36000)