diff --git a/src/squlearn/qnn/base_qnn.py b/src/squlearn/qnn/base_qnn.py
index fb763338..6f68ae6f 100644
--- a/src/squlearn/qnn/base_qnn.py
+++ b/src/squlearn/qnn/base_qnn.py
@@ -193,37 +193,7 @@ def fit(self, X, y, weights: np.ndarray = None) -> None:
                 Labels
             weights: Weights for each data point
         """
-        self.encoding_circuit._check_feature_consistency(X)
-        num_features = extract_num_features(X)
-
         self._is_lowlevel_qnn_initialized = False
-        self._initialize_lowlevel_qnn(num_features)
-
-        if self.param_ini is None or len(self.param_ini) != self._qnn.num_parameters:
-            self._param = self.encoding_circuit.generate_initial_parameters(
-                seed=self.parameter_seed, num_features=num_features
-            )
-        else:
-            self._param = self.param_ini.copy()
-
-        if (
-            self.param_op_ini is None
-            or len(self.param_op_ini) != self._qnn.num_parameters_observable
-        ):
-            if isinstance(self.operator, list):
-                self._param_op = np.concatenate(
-                    [
-                        operator.generate_initial_parameters(seed=self.parameter_seed + i + 1)
-                        for i, operator in enumerate(self.operator)
-                    ]
-                )
-            else:
-                self._param_op = self.operator.generate_initial_parameters(
-                    seed=self.parameter_seed + 1
-                )
-        else:
-            self._param_op = self.param_op_ini.copy()
-
         self._is_fitted = False
         self._fit(X, y, weights)
 
@@ -387,6 +357,35 @@ def _initialize_lowlevel_qnn(self, num_features: int | None = None) -> None:
             caching=self.caching,
             primitive=self.primitive,
         )
+
+        if self._is_fitted:
+            self._is_lowlevel_qnn_initialized = True
+            return
+
+        if self.param_ini is None or len(self.param_ini) != self.encoding_circuit.num_parameters:
+            self._param = self.encoding_circuit.generate_initial_parameters(
+                seed=self.parameter_seed, num_features=num_features
+            )
+        else:
+            self._param = self.param_ini.copy()
+
+        if (
+            self.param_op_ini is None
+            or len(self.param_op_ini) != self._qnn.num_parameters_observable
+        ):
+            if isinstance(self.operator, list):
+                self._param_op = np.concatenate(
+                    [
+                        operator.generate_initial_parameters(seed=self.parameter_seed + i + 1)
+                        for i, operator in enumerate(self.operator)
+                    ]
+                )
+            else:
+                self._param_op = self.operator.generate_initial_parameters(
+                    seed=self.parameter_seed + 1
+                )
+        else:
+            self._param_op = self.param_op_ini.copy()
         self._is_lowlevel_qnn_initialized = True
 
     def _validate_input(self, X, y, incremental, reset):