diff --git a/.github/workflows/samvg-cuda.yml b/.github/workflows/samvg-cuda.yml
new file mode 100644
index 0000000..32ff02e
--- /dev/null
+++ b/.github/workflows/samvg-cuda.yml
@@ -0,0 +1,42 @@
+name: SAMVG CUDA
+
+on:
+  schedule:
+    - cron: "17 3 * * 1"
+  workflow_dispatch:
+
+jobs:
+  extension:
+    name: Build and test CUDA renderer
+    runs-on: [self-hosted, linux, gpu]
+    timeout-minutes: 30
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: astral-sh/setup-uv@v3
+        with:
+          enable-cache: true
+
+      - name: Set up Python
+        run: uv python install 3.13
+
+      - name: Install CUDA-capable test dependencies
+        run: uv sync --extra dev --extra samvg
+
+      - name: Build the optional extension
+        run: VECTRIFY_BUILD_SAMVG_CUDA=1 uv build --wheel --no-build-isolation
+
+      - name: Install the built wheel
+        run: uv pip install --reinstall --no-deps dist/*.whl
+
+      - name: Exercise the CUDA renderer tests
+        run: uv run pytest tests/refine/test_filled_paths.py -q
+
+      - name: Benchmark one deterministic renderer step
+        run: uv run python scripts/check_cuda_renderer.py
+
+      - name: Preserve wheels for renderer benchmarking
+        uses: actions/upload-artifact@v4
+        with:
+          name: samvg-cuda-wheel
+          path: dist/*.whl
diff --git a/scripts/check_cuda_renderer.py b/scripts/check_cuda_renderer.py
new file mode 100644
index 0000000..8731cc7
--- /dev/null
+++ b/scripts/check_cuda_renderer.py
@@ -0,0 +1,34 @@
+"""Run one reproducible CUDA renderer timing without downloading a model."""
+
+from __future__ import annotations
+
+from time import perf_counter
+
+from PIL import Image
+
+from vectrify.refine import cuda_renderer
+from vectrify.refine.paths import fit_filled_svg
+
+SVG = """<svg xmlns=\"http://www.w3.org/2000/svg\" width=\"32\" height=\"32\">
+<path d=\"M 2 2 C 20 0 32 12 30 30 C 10 32 0 20 2 2 Z\" fill=\"#4080c0\"/>
+</svg>"""
+
+
+def main() -> None:
+    import torch
+
+    if not torch.cuda.is_available():
+        raise RuntimeError("SAMVG CUDA benchmark requires a GPU runner")
+    if cuda_renderer._extension() is None:
+        raise RuntimeError("SAMVG CUDA extension was not built")
+    target = Image.new("RGB", (32, 32), "#4080c0")
+    fit_filled_svg(SVG, target, steps=1, optimisation_long_side=32)
+    torch.cuda.synchronize()
+    started = perf_counter()
+    fit_filled_svg(SVG, target, steps=1, optimisation_long_side=32)
+    torch.cuda.synchronize()
+    print(f"samvg-cuda renderer step: {perf_counter() - started:.4f}s")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tests/refine/test_samvg_regression.py b/tests/refine/test_samvg_regression.py
new file mode 100644
index 0000000..9c8d21a
--- /dev/null
+++ b/tests/refine/test_samvg_regression.py
@@ -0,0 +1,51 @@
+"""Deterministic, model-free SAMVG seed regression coverage."""
+
+from __future__ import annotations
+
+import xml.etree.ElementTree as ET
+
+import numpy as np
+import pytest
+from PIL import Image
+
+from vectrify.refine.samvg import generate_svg
+
+
+@pytest.fixture
+def two_band_target() -> tuple[Image.Image, list[np.ndarray]]:
+    """A fixed seed image and its segmentation masks, with no model download."""
+    pixels = np.array(
+        [
+            [[20, 30, 40]] * 6,
+            [[20, 30, 40]] * 6,
+            [[220, 210, 200]] * 6,
+            [[220, 210, 200]] * 6,
+        ],
+        dtype=np.uint8,
+    )
+    upper = np.zeros((4, 6), dtype=bool)
+    upper[:2] = True
+    lower = np.zeros((4, 6), dtype=bool)
+    lower[2:] = True
+    return Image.fromarray(pixels), [upper, lower]
+
+
+def test_fixed_seed_image_exports_two_editable_coloured_paths(two_band_target):
+    image, masks = two_band_target
+
+    svg = generate_svg(
+        image,
+        masks,
+        min_pixels=1,
+        min_impact=0,
+        segments=4,
+        hybrid_strokes=False,
+        ocr=False,
+    )
+
+    root = ET.fromstring(svg)
+    paths = list(root)
+    assert root.attrib["viewBox"] == "0 0 6 4"
+    assert [path.attrib["fill"] for path in paths] == ["#141e28", "#dcd2c8"]
+    assert all(path.attrib["fill-rule"] == "evenodd" for path in paths)
+    assert all(path.attrib["d"].endswith(" Z") for path in paths)