diff --git a/mohou/encoder.py b/mohou/encoder.py
index bc54b60..3589369 100644
--- a/mohou/encoder.py
+++ b/mohou/encoder.py
@@ -13,7 +13,15 @@
 from mohou.model.autoencoder import AutoEncoderBase
 from mohou.model.common import ModelBase
 from mohou.trainer import TrainCache
-from mohou.types import ElementT, EpisodeBundle, ImageT, VectorT, get_element_type
+from mohou.types import (
+    ElementT,
+    EpisodeBundle,
+    ImageT,
+    PrimitiveImageBase,
+    PrimitiveImageT,
+    VectorT,
+    get_element_type,
+)
 from mohou.utils import assert_equal_with_message, assert_isinstance_with_message
 
 logger = logging.getLogger(__name__)
@@ -96,6 +104,54 @@ def set_device(self, device: torch.device) -> None:
         self._get_model().put_on_device(device)
 
 
+@dataclass
+class PCAImageEncoder(EncoderBase[PrimitiveImageT]):
+    pca: PCA
+
+    @classmethod
+    def from_bundle(
+        cls, bundle: EpisodeBundle, image_type: Type[PrimitiveImageT], n_out: int
+    ) -> "PCAImageEncoder":
+        assert issubclass(image_type, PrimitiveImageBase), "currently only support PrimitiveImage"
+        vector_list = []
+        image_shape = None
+        for episode in bundle.get_touch_bundle():
+            image_seq = episode.get_sequence_by_type(image_type)
+            vecs = [image.numpy().flatten() for image in image_seq.elem_list]
+            vector_list.extend(vecs)
+            if image_shape is None:
+                image_shape = image_seq.elem_list[0].shape
+            else:
+                assert image_shape == image_seq.elem_list[0].shape
+        assert image_shape is not None
+        mat = np.array(vector_list)
+        pca = PCA(n_components=n_out)
+        pca.fit(mat)
+        return cls(image_type, image_shape, n_out, pca)
+
+    def _forward_impl(self, inp: ImageT) -> np.ndarray:
+        assert isinstance(inp, PrimitiveImageBase)
+        inp_as_2d = inp.numpy().flatten().reshape(1, -1)
+        out = self.pca.transform(inp_as_2d)
+        return out.flatten()
+
+    def _backward_impl(self, inp: np.ndarray) -> PrimitiveImageT:
+        out = self.pca.inverse_transform(inp.reshape(1, -1))
+        out_reshaped = out.reshape(self.input_shape)
+        out_uint8 = out_reshaped.astype(np.uint8)
+        return self.elem_type(out_uint8)
+
+    def save(self, project_path: Path) -> None:
+        with (project_path / "image_pca.pkl").open("wb") as f:
+            pickle.dump(self, f)
+
+    @classmethod
+    def load(cls, project_path: Path) -> "PCAImageEncoder":
+        with (project_path / "image_pca.pkl").open("rb") as f:
+            pca = pickle.load(f)
+        return pca
+
+
 @dataclass(eq=False)
 class ImageEncoder(EncoderBase[ImageT], HasAModel):
     input_shape: Tuple[int, int, int]