diff --git a/lato_mlx/models/vvae.py b/lato_mlx/models/vvae.py
index 5065786..02f2335 100644
--- a/lato_mlx/models/vvae.py
+++ b/lato_mlx/models/vvae.py
@@ -46,7 +46,10 @@ class DownResBlock(nn.Module):
         self.block = SparseResBlock(channels, out_channels)
 
     def __call__(self, x: SparseTensor) -> SparseTensor:
-        return self.block(downsample(x, 2))
+        # mode="max" explicitly: LATO.2 hardcodes reduce="amax", but Pixal3D's version of
+        # the same op defaults to "mean". Relying on the shared default would silently
+        # change every downsampled feature the day that default moves.
+        return self.block(downsample(x, 2, mode="max"))
 
 
 class SparseTransformerBase(nn.Module):