diff --git a/src/spatialdata/models/models.py b/src/spatialdata/models/models.py
index 01b34e76..92acbec4 100644
--- a/src/spatialdata/models/models.py
+++ b/src/spatialdata/models/models.py
@@ -891,7 +891,7 @@ def _add_metadata_and_validate(
             # It also just changes the state of the series, so it is not a big deal.
             if isinstance(data[c].dtype, CategoricalDtype) and not data[c].cat.known:
                 try:
-                    data[c] = data[c].cat.set_categories(data[c].head(1).cat.categories)
+                    data[c] = data[c].cat.set_categories(data[c].compute().cat.categories)
                 except ValueError:
                     logger.info(f"Column `{c}` contains unknown categories. Consider casting it.")
 
diff --git a/tests/models/test_models.py b/tests/models/test_models.py
index 124933f4..1e82b698 100644
--- a/tests/models/test_models.py
+++ b/tests/models/test_models.py
@@ -830,3 +830,31 @@ def test_warning_on_large_chunks():
         assert len(w) == 1, "Warning should be raised for large chunk size"
         assert issubclass(w[-1].category, UserWarning)
         assert "Detected chunks larger than:" in str(w[-1].message)
+
+
+def test_categories_on_partitioned_dataframe(sdata_blobs: SpatialData):
+    df = sdata_blobs["blobs_points"].compute()
+    df["genes"] = RNG.choice([f"gene_{i}" for i in range(200)], len(df))
+    N_PARTITIONS = 200
+    ddf = dd.from_pandas(df, npartitions=N_PARTITIONS)
+    ddf["genes"] = ddf["genes"].astype("category")
+
+    df["genes"] = df["genes"].astype("category")
+    df_parsed = PointsModel.parse(df, npartitions=N_PARTITIONS)
+    ddf_parsed = PointsModel.parse(ddf, npartitions=N_PARTITIONS)
+
+    assert df["genes"].equals(df_parsed["genes"].compute())
+    assert df["genes"].cat.categories.equals(df_parsed["genes"].compute().cat.categories)
+
+    assert np.array_equal(df["genes"].to_numpy(), ddf_parsed["genes"].compute().to_numpy())
+    assert set(df["genes"].cat.categories.tolist()) == set(ddf_parsed["genes"].compute().cat.categories.tolist())
+
+    # two behavior to investigate later/report to dask (they originate in dask)
+    # TODO: df['genes'].cat.categories has dtype 'object', while ddf_parsed['genes'].compute().cat.categories has dtype
+    #  'string'
+    # this problem should disappear after pandas 3.0 is released
+    assert df["genes"].cat.categories.dtype == "object"
+    assert ddf_parsed["genes"].compute().cat.categories.dtype == "string"
+
+    # TODO: the list of categories are not preserving the order
+    assert df["genes"].cat.categories.tolist() != ddf_parsed["genes"].compute().cat.categories.tolist()