diff --git a/kauldron/modules/__init__.py b/kauldron/modules/__init__.py
index 805f1b49..4068fd9b 100644
--- a/kauldron/modules/__init__.py
+++ b/kauldron/modules/__init__.py
@@ -42,3 +42,4 @@
 from kauldron.modules.pos_embeddings import FourierEmbedding
 from kauldron.modules.pos_embeddings import LearnedEmbedding
 from kauldron.modules.pos_embeddings import ZeroEmbedding
+from kauldron.modules.pos_embeddings import Add2DRoPE
diff --git a/kauldron/modules/pos_embeddings.py b/kauldron/modules/pos_embeddings.py
index 7a64bf47..dc9cf8f0 100644
--- a/kauldron/modules/pos_embeddings.py
+++ b/kauldron/modules/pos_embeddings.py
@@ -24,6 +24,42 @@
 from kauldron.typing import Axes, Dtype, Float, Initializer, Shape, typechecked  # pylint: disable=g-multiple-import,g-importing-member
 
 
+class Add2DRoPE(nn.Module):
+  """Helper Module for adding a 2D version of RoPE embedding to inputs.
+
+  Attributes:
+    num_wavelengths: Number of wavelengths.
+  """
+
+  num_wavelengths: int
+
+  def __call__(self, inputs: Float['*B X Y C']) -> Float['*B X Y C']:
+    num_wavelengths = inputs.shape[-1]
+    assert num_wavelengths % 4 == 0
+    num_channels = num_wavelengths // 2
+    freq_exponents = (2. / num_channels) * jnp.arange(num_channels // 2)
+    timescale = (num_wavelengths ** freq_exponents)[-3]
+    posx, posy = jnp.arange(inputs.shape[-3]), jnp.arange(inputs.shape[-2])
+    radx = posx[None, None, ..., None] / timescale[None, None, None, :]
+    # radians.shape = [...,L,1,d=D/2]
+    sinx, cosx = jnp.sin(radx), jnp.cos(radx)
+    rady = posy[None, None, ..., None] / timescale[None, None, None, :]
+    # radians.shape = [...,L,1,d=D/2]
+    siny, cosy = jnp.sin(rady), jnp.cos(rady)
+
+    x_inputs, y_inputs = jnp.split(inputs, 2, axis=-1)
+    x1, x2 = jnp.split(x_inputs, 2, axis=-1)
+    cosx, sinx = jnp.expand_dims(cosx, axis=-2), jnp.expand_dims(sinx, axis=-2)
+    cosy, siny = jnp.expand_dims(cosy, axis=-3), jnp.expand_dims(siny, axis=-3)
+    x_res = jnp.concatenate(
+        [x1 * cosx - x2 * sinx, x2 * cosx + x1 * sinx], axis=-1)
+    y1, y2 = jnp.split(y_inputs, 2, axis=-1)
+    y_res = jnp.concatenate(
+        [y1 * cosy - y2 * siny, y2 * cosy + y1 * siny], axis=-1)
+    res = jnp.concatenate([x_res, y_res], axis=-1)
+    return res
+
+
 class AddEmbedding(nn.Module):
   """Helper Module for adding a PositionEmbedding e.g. in a `knn.Sequential`.