diff --git a/README.md b/README.md
index 5789fb0..404211c 100644
--- a/README.md
+++ b/README.md
@@ -104,7 +104,7 @@ If you face out of memory issues, you can reduce the `num_views` or the `batch_s
 
 ```yaml
 train:
-  batch_size: 1                     # Batch size
+  batch_size: 1
   camera:
     num_views: 6  
 ```
diff --git a/utils/render.py b/utils/render.py
index 48a5cbe..bdbaf29 100644
--- a/utils/render.py
+++ b/utils/render.py
@@ -123,6 +123,7 @@ def to_pil(rendered_features, target_channels=(0, 3),
 
         :return: A PIL Image showing the rendered images.
         """
+        # @ TODO This function is not consistent with the meshnca.render_channels
         assert rendered_features.dim() == 5, "The input tensor should have 5 dimensions"
 
         if not isinstance(target_channels, dict):
@@ -143,7 +144,7 @@ def to_pil(rendered_features, target_channels=(0, 3),
         )  # [batch_size*height, num_views*width, num_features]
 
         image_list = []
-        for key, channels in target_channels.items():
+        for key, channels in sorted(target_channels.items()):
             image = features[..., range(*channels)]
             if image.shape[-1] == 1:
                 image = np.repeat(image, 3, axis=-1)