diff --git a/requirements.txt b/requirements.txt
index fc4f51a..b763743 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,2 +1,4 @@
+numpy==1.26.4
 numcodecs
 blosc2
+pytest
diff --git a/tests/test_ocf_blosc2.py b/tests/test_ocf_blosc2.py
index 5b0f2df..bf80d04 100644
--- a/tests/test_ocf_blosc2.py
+++ b/tests/test_ocf_blosc2.py
@@ -1,111 +1,14 @@
-"""Unit tests for satip.jpeg_xl_float_with_nans."""
-
-import unittest
+"""Unit tests for ocf_blosc2"""
 
 import numpy as np
 
 import ocf_blosc2
 
 
-class TestJpegXlFloatWithNaNs(unittest.TestCase):
-    """Test class for unittest for the class methods and the functions.
-
-    We only test our home-written functions.
-    The two similarly named class functions encode and decode are mostly wrappers
-    around our home-written function output piped into an external library.
-    Testing the functionality of the external functions is out of scope.
-    """
-
-    def setUp(self) -> None:  # noqa D102
-        # Define synthetic input array and the expected target array:
-        self.buf = np.asarray([np.nan, 0.0, 0.5, 1.0], dtype=np.float32)
-        self.encoded = np.asarray(
-            [NAN_VALUE, LOWER_BOUND_FOR_REAL_PIXELS, 0.5 * (1 + LOWER_BOUND_FOR_REAL_PIXELS), 1.0],
-            dtype=np.float32,
-        )
-
-        self.jpegxl = ocf_blosc2.Blosc2()
-
-        return super().setUp()
-
-    def test_encode(self):
-        """Tests the encoding function.
-
-        After encoding the raw array, the nan-values should be gone and the
-        real values should be transformed to the range specified by the
-        constants imported from the source code. See there for more details.
-        """
-        # Check that the enconded buffer matches the expected target
-        # (attention: use a copy of the originals!):
-        self.assertTrue(np.isclose(encode_nans(self.buf.copy()), self.encoded).all())
-
-    def test_decode(self):
-        """Tests the decoding function.
-
-        When taking what was previously the encoded array and decode it,
-        we expect to get the original buf-array back again.
-        """
-        # As np.nan != np.nan (!) and thus np.isclose or array comparison do not consider
-        # two nan-values to be close or equal, we have to replace all nan-values with
-        # a numeric value before comparison. This numeric value should be one that
-        # can not be created via decoding (e.g. a negative number).
-        nan_replacement = -3.14
-        self.assertTrue(
-            np.isclose(
-                np.nan_to_num(self.buf, nan_replacement),
-                np.nan_to_num(decode_nans(self.encoded.copy()), nan_replacement),
-            ).all()
-        )
-
-    def test_class_roundtrip(self):
-        """Tests the class-defined wrappers around our home-written functions.
-
-        We test whether a back-and-forth transformation (nested encode-decode)
-        will give us back our original input value.
-        """
-        reshaped_buf = self.buf.copy().reshape((1, -1, 1))
-
-        roundtrip_result = self.jpegxl.decode(self.jpegxl.encode(reshaped_buf.copy()))
-
-        # For reasons explained in the decoding test, we have to manually replace
-        # the nan-values to make them comparable:
-        nan_replacement = -3.14
-        reshaped_buf = np.nan_to_num(reshaped_buf, nan_replacement)
-        roundtrip_result = np.nan_to_num(roundtrip_result, nan_replacement)
-
-        # When we do the comparison, we have to be very lenient, as the external library
-        # will have worked its compression magic, so values will not completely align.
-        # Also, going back and forth removes the information about the channel number
-        # in our test case (presumably b/c we here only have one channel for simplicity's sake).
-        # So we have to reshape both:
-        self.assertTrue(
-            np.isclose(reshaped_buf.reshape((-1)), roundtrip_result.reshape((-1)), atol=0.1).all()
-        )
-
-    def test_consistent_init_params(self):
-        """The JpegXLFloat-class has to be initialised with specific parameter combinations.
-
-        Stuff that is allowed:
-        1. If lossless = None, then everything is allowed.
-        2. If lossless = True, then level has to be None and distance has to be None or 0
-        3. If lossless = False, then everything is allowed.
-
-        To test this, we will try various parameters and see that the class gets
-        initialised properly, w/o throwing any errors.
-        """
-
-        # Sub-case 1:
-        self.assertTrue(JpegXlFloatWithNaNs(lossless=None, level="very_high", distance=-10))
-
-        # Sub-case 2:
-        with self.assertRaises(AssertionError):
-            JpegXlFloatWithNaNs(lossless=True, level=1, distance=1)
-
-        with self.assertRaises(AssertionError):
-            JpegXlFloatWithNaNs(lossless=True, level=None, distance=1)
-
-        with self.assertRaises(AssertionError):
-            JpegXlFloatWithNaNs(lossless=True, level=2, distance=0)
-
-        # Sub-case 3:
-        self.assertTrue(JpegXlFloatWithNaNs(lossless=False))
+def test_roundtrip():
+    buf = np.asarray([np.nan, 0.0, 0.5, 1.0], dtype=np.float32)
+    blosc2 = ocf_blosc2.Blosc2()
+    comp_arr = blosc2.encode(buf)
+    dest = np.empty(buf.shape, buf.dtype)
+    blosc2.decode(comp_arr, out=dest)
+    assert np.allclose(buf, dest, equal_nan=True)