diff --git a/tests/test_frame_selector.py b/tests/test_frame_selector.py
new file mode 100644
index 0000000..6cd1404
--- /dev/null
+++ b/tests/test_frame_selector.py
@@ -0,0 +1,113 @@
+import importlib.util
+from pathlib import Path
+
+import cv2
+import numpy as np
+import torch
+
+
+MODULE_PATH = Path(__file__).resolve().parents[1] / "frame_selector.py"
+SPEC = importlib.util.spec_from_file_location("frame_selector", MODULE_PATH)
+frame_selector = importlib.util.module_from_spec(SPEC)
+SPEC.loader.exec_module(frame_selector)
+
+AIA_SelectVideoFrame = frame_selector.AIA_SelectVideoFrame
+_normalize_selector = frame_selector._normalize_selector
+
+
+def _checker(size=64, block=8):
+    y, x = np.indices((size, size))
+    pattern = ((x // block + y // block) % 2) * 255
+    return np.stack([pattern, pattern, pattern], axis=-1).astype(np.uint8)
+
+
+def _to_tensor(frames):
+    return torch.from_numpy(np.stack(frames).astype(np.float32) / 255.0)
+
+
+def _disable_face_detection(monkeypatch, count):
+    monkeypatch.setattr(
+        AIA_SelectVideoFrame,
+        "_detect_face_scores",
+        staticmethod(lambda _frames: np.zeros(count, dtype=np.float32)),
+    )
+
+
+def test_selector_aliases():
+    assert _normalize_selector("") == "best"
+    assert _normalize_selector("best still") == "best"
+    assert _normalize_selector("no face please") == "no_face"
+    assert _normalize_selector("very blurry") == "blurry"
+    assert _normalize_selector("moving") == "moving"
+    assert _normalize_selector("static pose") == "static"
+    assert _normalize_selector("strong lighting flicker") == "lighting"
+    assert _normalize_selector("whatever") == "best"
+
+
+def test_blurry_mode_selects_soft_frame(monkeypatch):
+    sharp = _checker()
+    soft = cv2.GaussianBlur(sharp, (17, 17), 0)
+    frames = [sharp, soft, sharp]
+    _disable_face_detection(monkeypatch, len(frames))
+
+    selected = AIA_SelectVideoFrame().select(_to_tensor(frames), "blurry")[0]
+
+    assert torch.allclose(selected[0], _to_tensor([soft])[0])
+
+
+def test_moving_and_static_modes_split_on_motion(monkeypatch):
+    base = _checker()
+    shifted = np.roll(base, 18, axis=1)
+    frames = [base, base, base, shifted, base, base, base]
+    _disable_face_detection(monkeypatch, len(frames))
+    node = AIA_SelectVideoFrame()
+
+    moving = node.select(_to_tensor(frames), "moving")[0]
+    static = node.select(_to_tensor(frames), "static")[0]
+
+    assert torch.allclose(moving[0], _to_tensor([shifted])[0])
+    assert torch.allclose(static[0], _to_tensor([base])[0])
+
+
+def test_lighting_mode_selects_brightness_jump(monkeypatch):
+    dark = np.full((64, 64, 3), 40, dtype=np.uint8)
+    pre_jump = np.full((64, 64, 3), 80, dtype=np.uint8)
+    bright = np.full((64, 64, 3), 220, dtype=np.uint8)
+    post_jump = np.full((64, 64, 3), 90, dtype=np.uint8)
+    frames = [dark.copy() for _ in range(31)]
+    frames[15] = pre_jump
+    frames[16] = bright
+    frames[17] = post_jump
+    _disable_face_detection(monkeypatch, len(frames))
+
+    selected = AIA_SelectVideoFrame().select(_to_tensor(frames), "flicker")[0]
+
+    selected_mean = int(round(float(selected[0].mean() * 255)))
+    assert selected_mean in {80, 90, 220}
+
+
+def test_no_face_mode_uses_face_scores(monkeypatch):
+    frames = [_checker() for _ in range(5)]
+    face_scores = np.asarray([1.0, 0.8, 0.0, 0.7, 1.0], dtype=np.float32)
+    monkeypatch.setattr(
+        AIA_SelectVideoFrame,
+        "_detect_face_scores",
+        staticmethod(lambda _frames: face_scores),
+    )
+
+    selected = AIA_SelectVideoFrame().select(_to_tensor(frames), "no face")[0]
+
+    assert torch.allclose(selected[0], _to_tensor([frames[2]])[0])
+
+
+def test_smoke_241_frame_batch(monkeypatch):
+    frames = []
+    base = _checker()
+    for idx in range(241):
+        frame = np.roll(base, idx % 32, axis=1)
+        frames.append(frame)
+    _disable_face_detection(monkeypatch, len(frames))
+
+    selected = AIA_SelectVideoFrame().select(_to_tensor(frames), "")[0]
+
+    assert selected.shape == (1, 64, 64, 3)