diff --git a/src/h3_blackwell_runtime/vae_encoder.py b/src/h3_blackwell_runtime/vae_encoder.py
index e6dcd5d..29d0d64 100644
--- a/src/h3_blackwell_runtime/vae_encoder.py
+++ b/src/h3_blackwell_runtime/vae_encoder.py
@@ -56,11 +56,15 @@ def _causal_conv3d(x, weight, bias, *, kernel_size, stride, spatial_padding):
     #   x = F.pad(x, (0, 0, 0, 0, (k-1), 0), mode="constant")     -- temporal
     #   return super().forward(x)
     if x.shape[2] == 1:
-        # Keyframe path (matches reference's `autopad="causal_zero"`): truncate
-        # the temporal taps to the center 1 and run as an effective 2D conv.
+        # Keyframe path (matches reference's `autopad="causal_zero"`): apply
+        # spatial-reflect pad, then truncate the temporal taps to the center 1
+        # so the conv acts as a 2D op on the single frame.
+        if spatial_padding > 0:
+            # F.pad for 5D input (B,C,T,H,W) uses (W_l, W_r, H_l, H_r, T_l, T_r).
+            x = F.pad(x, (spatial_padding, spatial_padding, spatial_padding, spatial_padding, 0, 0), mode="reflect")
         half = (kernel_size - 1) // 2
         kernel_5d = weight[:, :, half:half + 1, :, :]
-        return F.conv3d(x, kernel_5d, bias, (1, stride[1], stride[2]), (spatial_padding, spatial_padding, spatial_padding))
+        return F.conv3d(x, kernel_5d, bias, (1, stride[1], stride[2]), (0, 0, 0))
     if spatial_padding > 0:
         return _reflect_and_causal(x, weight, bias, kernel_size, stride, spatial_padding)
     return F.conv3d(x, weight, bias, stride, (0, 0, 0))