diff --git a/gemma/gm/nn/vision/_image.py b/gemma/gm/nn/vision/_image.py
index ce4182e0..ebbf39dc 100644
--- a/gemma/gm/nn/vision/_image.py
+++ b/gemma/gm/nn/vision/_image.py
@@ -16,6 +16,8 @@
 
 from __future__ import annotations
 from collections.abc import Sequence
+import io
+
 import einops
 from etils import epath
 import jax
@@ -23,7 +25,6 @@
 from kauldron import typing
 import numpy as np
 from PIL import Image
-import tensorflow as tf
 
 _IMAGE_MEAN = (127.5,) * 3
 _IMAGE_STD = (127.5,) * 3
@@ -69,11 +70,13 @@ def pre_process_image(
   Returns:
     The pre-processed image.
   """
-  # all inputs are expected to have been jpeg compressed.
-  # TODO(eyvinec): we should remove tf dependency.
-  image = jnp.asarray(
-      tf.image.decode_jpeg(tf.io.encode_jpeg(image), channels=3)
-  )
+  # All inputs are expected to have been JPEG-compressed.  Simulate the
+  # lossy round-trip so pixel values match what the model saw during training.
+  image_uint8 = np.asarray(image, dtype=np.uint8)
+  buf = io.BytesIO()
+  Image.fromarray(image_uint8).save(buf, format="JPEG")
+  buf.seek(0)
+  image = jnp.asarray(np.array(Image.open(buf).convert("RGB")))
   image = jax.image.resize(
       image,
       shape=(image_height, image_width, 3),