diff --git a/server/asset_generator.py b/server/asset_generator.py
new file mode 100644
index 0000000..639e542
--- /dev/null
+++ b/server/asset_generator.py
@@ -0,0 +1,57 @@
+"""SDXL Turbo wrapper for horror image generation."""
+
+import io
+
+import torch
+
+try:
+    from diffusers import AutoPipelineForText2Image
+except ImportError:  # pragma: no cover - exercised only in test environments
+    AutoPipelineForText2Image = None  # Tests patch this attribute directly.
+
+from server.config import config
+from server.prompts import NEGATIVE_PROMPT
+
+
+class AssetGenerator:
+    """Generates horror images via SDXL Turbo."""
+
+    def __init__(self, device: str | None = None, model_id: str | None = None):
+        self.device = device or config.device
+        self.model_id = model_id or config.models.sdxl_model_id
+
+        if AutoPipelineForText2Image is None:
+            raise RuntimeError(
+                "diffusers is not installed; install diffusers to use AssetGenerator"
+            )
+
+        use_fp16 = self.device == "cuda"
+        self._pipe = AutoPipelineForText2Image.from_pretrained(
+            self.model_id,
+            torch_dtype=torch.float16 if use_fp16 else torch.float32,
+            variant="fp16" if use_fp16 else None,
+        )
+        if self.device == "cuda":
+            self._pipe = self._pipe.to("cuda")
+
+    def generate(self, prompt: str, seed: int | None = None) -> bytes:
+        """Generate a 512x512 PNG image from a horror prompt. Returns PNG bytes."""
+        if seed is None:
+            seed = torch.randint(0, 2**32, (1,)).item()
+
+        generator = torch.Generator(device=self.device).manual_seed(seed)
+
+        result = self._pipe(
+            prompt=prompt,
+            negative_prompt=NEGATIVE_PROMPT,
+            num_inference_steps=config.models.sdxl_steps,
+            guidance_scale=config.models.sdxl_guidance_scale,
+            width=config.models.sdxl_width,
+            height=config.models.sdxl_height,
+            generator=generator,
+        )
+
+        image = result.images[0]
+        buf = io.BytesIO()
+        image.save(buf, format="PNG")
+        return buf.getvalue()
diff --git a/tests/test_asset_generator.py b/tests/test_asset_generator.py
new file mode 100644
index 0000000..050d766
--- /dev/null
+++ b/tests/test_asset_generator.py
@@ -0,0 +1,52 @@
+from unittest.mock import MagicMock, patch
+from PIL import Image
+
+from server.asset_generator import AssetGenerator
+
+
+class TestAssetGenerator:
+    @patch("server.asset_generator.AutoPipelineForText2Image")
+    def test_init_loads_model(self, mock_pipeline_cls):
+        """Generator loads the SDXL Turbo pipeline on init."""
+        mock_pipe = MagicMock()
+        mock_pipeline_cls.from_pretrained.return_value = mock_pipe
+        mock_pipe.to.return_value = mock_pipe
+
+        gen = AssetGenerator(device="cpu")
+        mock_pipeline_cls.from_pretrained.assert_called_once()
+
+    @patch("server.asset_generator.AutoPipelineForText2Image")
+    def test_generate_returns_bytes(self, mock_pipeline_cls):
+        """Generate returns PNG bytes."""
+        mock_pipe = MagicMock()
+        mock_pipeline_cls.from_pretrained.return_value = mock_pipe
+        mock_pipe.to.return_value = mock_pipe
+
+        # Mock pipeline output
+        fake_image = Image.new("RGB", (512, 512), color="black")
+        mock_result = MagicMock()
+        mock_result.images = [fake_image]
+        mock_pipe.return_value = mock_result
+
+        gen = AssetGenerator(device="cpu")
+        data = gen.generate("dark void, horror")
+        assert isinstance(data, bytes)
+        assert len(data) > 0
+
+    @patch("server.asset_generator.AutoPipelineForText2Image")
+    def test_generate_uses_negative_prompt(self, mock_pipeline_cls):
+        """Generate passes the negative prompt to the pipeline."""
+        mock_pipe = MagicMock()
+        mock_pipeline_cls.from_pretrained.return_value = mock_pipe
+        mock_pipe.to.return_value = mock_pipe
+
+        fake_image = Image.new("RGB", (512, 512), color="black")
+        mock_result = MagicMock()
+        mock_result.images = [fake_image]
+        mock_pipe.return_value = mock_result
+
+        gen = AssetGenerator(device="cpu")
+        gen.generate("test prompt")
+
+        call_kwargs = mock_pipe.call_args
+        assert "negative_prompt" in call_kwargs.kwargs