diff --git a/gemma/gm/nn/_modules_test.py b/gemma/gm/nn/_modules_test.py
index afde3dcf..c81c0eed 100644
--- a/gemma/gm/nn/_modules_test.py
+++ b/gemma/gm/nn/_modules_test.py
@@ -71,7 +71,57 @@ def test_embedder_decode():
   np.testing.assert_array_equal(output, jnp.array(expected))
 
 
-# TODO(mblondel): Add tests for `encode_vision` here.
+def test_encode_vision_output_shape():
+  """encode_vision should project from vision_proj_dim to embed_dim."""
+  vocab_size = 10
+  embed_dim = 8
+  vision_proj_dim = 16
+  embedder = gm.nn.Embedder(
+      vocab_size=vocab_size,
+      embed_dim=embed_dim,
+      vision_proj_dim=vision_proj_dim,
+  )
+
+  rng = jax.random.PRNGKey(0)
+  dummy_vision_input = jnp.ones((1, 4, vision_proj_dim))
+  params = embedder.init(rng, dummy_vision_input, method=embedder.encode_vision)
+
+  # [B, num_patches, vision_proj_dim] -> [B, num_patches, embed_dim]
+  vision_input = jax.random.normal(rng, (2, 4, vision_proj_dim))
+  output = embedder.apply(
+      params, vision_input, method=embedder.encode_vision
+  )
+  assert output.shape == (2, 4, embed_dim)
+
+
+def test_encode_vision_different_batch_shapes():
+  """encode_vision should handle various batch/patch dimensions."""
+  vocab_size = 10
+  embed_dim = 8
+  vision_proj_dim = 16
+  embedder = gm.nn.Embedder(
+      vocab_size=vocab_size,
+      embed_dim=embed_dim,
+      vision_proj_dim=vision_proj_dim,
+  )
+
+  rng = jax.random.PRNGKey(0)
+  dummy_input = jnp.ones((1, 1, vision_proj_dim))
+  params = embedder.init(rng, dummy_input, method=embedder.encode_vision)
+
+  output_1 = embedder.apply(
+      params,
+      jax.random.normal(rng, (1, 1, vision_proj_dim)),
+      method=embedder.encode_vision,
+  )
+  assert output_1.shape == (1, 1, embed_dim)
+
+  output_8 = embedder.apply(
+      params,
+      jax.random.normal(rng, (1, 8, vision_proj_dim)),
+      method=embedder.encode_vision,
+  )
+  assert output_8.shape == (1, 8, embed_dim)
 
 
 def test_sliding_mask():