diff --git a/bdpy/dl/torch/dataset.py b/bdpy/dl/torch/dataset.py
index c821592..c45f616 100644
--- a/bdpy/dl/torch/dataset.py
+++ b/bdpy/dl/torch/dataset.py
@@ -157,9 +157,16 @@ class ImageDataset(Dataset):
     root_path : str | Path
         Path to the root directory of images.
     stimulus_names : list[str], optional
-        List of stimulus names. If None, all stimulus names are used.
+        List of stimulus names. If provided, images are returned in the given
+        order. If None, all images found under ``root_path`` are used in
+        alphabetical order.
     extension : str, optional
         Extension of the image files.
+
+    Notes
+    -----
+    Images are returned as float64 arrays in CHW (channels, height, width)
+    format with pixel values normalized to [0, 1].
     """
 
     def __init__(
@@ -170,10 +177,7 @@ def __init__(
     ):
         self.root_path = root_path
         if stimulus_names is None:
-            stimulus_names = [
-                _removesuffix(path.name, "." + extension)
-                for path in Path(root_path).glob(f"*{extension}")
-            ]
+            stimulus_names = sorted(path.stem for path in Path(root_path).glob(f"*{extension}"))
         self._stimulus_names = stimulus_names
         self._extension = extension
 
@@ -184,7 +188,7 @@ def __getitem__(self, index: int):
         stimulus_name = self._stimulus_names[index]
         image = Image.open(Path(self.root_path) / f"{stimulus_name}.{self._extension}")
         image = image.convert("RGB")
-        return np.array(image) / 255.0, stimulus_name
+        return np.array(image).transpose(2, 0, 1) / 255.0, stimulus_name
 
 
 class RenameFeatureKeys:
diff --git a/tests/dl/torch/test_dataset.py b/tests/dl/torch/test_dataset.py
new file mode 100644
index 0000000..ff0be98
--- /dev/null
+++ b/tests/dl/torch/test_dataset.py
@@ -0,0 +1,89 @@
+import tempfile
+import unittest
+from pathlib import Path
+
+import numpy as np
+from PIL import Image
+from torch.utils.data import DataLoader
+
+from bdpy.dl.torch.dataset import ImageDataset
+
+# Use non-square images (H=4, W=6, C=3) to fully discriminate every axis.
+# A square image (H=W) cannot distinguish (3, H, H) from (H, H, 3).
+_H, _W = 4, 6
+
+
+def _save_image(path: Path, r: int, g: int, b: int, h: int = _H, w: int = _W) -> None:
+    data = np.zeros((h, w, 3), dtype=np.uint8)
+    data[:, :, 0] = r
+    data[:, :, 1] = g
+    data[:, :, 2] = b
+    Image.fromarray(data).save(path)
+
+
+class TestImageDataset(unittest.TestCase):
+    def setUp(self) -> None:
+        self.tmpdir = tempfile.TemporaryDirectory()
+        root = Path(self.tmpdir.name)
+        _save_image(root / "a.png", r=200, g=100, b=50)
+        _save_image(root / "b.png", r=10, g=20, b=30)
+        _save_image(root / "c.png", r=0, g=128, b=255)
+        self.root = root
+
+    def tearDown(self) -> None:
+        self.tmpdir.cleanup()
+
+    def test_getitem_returns_chw_shape(self):
+        dataset = ImageDataset(self.root, stimulus_names=["a"], extension="png")
+        arr, _ = dataset[0]
+        self.assertEqual(arr.shape, (3, _H, _W))
+
+    def test_getitem_preserves_channels(self):
+        # R=200 G=100 B=50 — verifies C axis maps to the correct channel.
+        dataset = ImageDataset(self.root, stimulus_names=["a"], extension="png")
+        arr, _ = dataset[0]
+        self.assertTrue(np.allclose(arr[0], 200 / 255.0))
+        self.assertTrue(np.allclose(arr[1], 100 / 255.0))
+        self.assertTrue(np.allclose(arr[2], 50 / 255.0))
+
+    def test_dataloader_integration_batch_shape(self):
+        dataset = ImageDataset(self.root, stimulus_names=["a", "b"], extension="png"    )
+        loader = DataLoader(dataset, batch_size=2)
+        batch_images, _ = next(iter(loader))
+        self.assertEqual(tuple(batch_images.shape), (2, 3, _H, _W))
+
+    def test_value_range_normalized_to_unit_interval(self):
+        dataset = ImageDataset(self.root, stimulus_names=["a", "b", "c"], extension="png")
+        for i in range(len(dataset)):
+            arr, _ = dataset[i]
+            self.assertGreaterEqual(float(arr.min()), 0.0)
+            self.assertLessEqual(float(arr.max()), 1.0)
+
+    def test_len_matches_stimulus_names(self):
+        dataset = ImageDataset(self.root, stimulus_names=["a", "b", "c"], extension="png")
+        self.assertEqual(len(dataset), 3)
+
+    def test_explicit_stimulus_names_respected(self):
+        dataset = ImageDataset(self.root, stimulus_names=["a", "c"], extension="png")
+        self.assertEqual(len(dataset), 2)
+        _, label0 = dataset[0]
+        _, label1 = dataset[1]
+        self.assertEqual(label0, "a")
+        self.assertEqual(label1, "c")
+
+    def test_auto_detected_stimulus_names_use_stem(self):
+        dataset = ImageDataset(self.root, extension="png")
+        self.assertEqual(set(dataset._stimulus_names), {"a", "b", "c"})
+
+    def test_explicit_stimulus_names_preserve_input_order(self):
+        dataset = ImageDataset(self.root, stimulus_names=["c", "a", "b"], extension="png")
+        labels = [dataset[i][1] for i in range(len(dataset))]
+        self.assertEqual(labels, ["c", "a", "b"])
+
+    def test_auto_detected_stimulus_names_are_sorted(self):
+        dataset = ImageDataset(self.root, extension="png")
+        self.assertEqual(dataset._stimulus_names, ["a", "b", "c"])
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/tests/dl/torch/test_torch.py b/tests/dl/torch/test_torch.py
index c5bcc2d..bcaf51e 100644
--- a/tests/dl/torch/test_torch.py
+++ b/tests/dl/torch/test_torch.py
@@ -69,9 +69,5 @@ def test_run_with_layer_map(self):
             self.assertEqual(features[layer].shape, shape)
 
 
-class TestImageDataset(unittest.TestCase):
-    ...
-
-
 if __name__ == '__main__':
     unittest.main()
\ No newline at end of file