diff --git a/hy3dshape/configs/hunyuandit-mini-overfitting-flowmatching-dinol518-bf16-lr1e4-4096.yaml b/hy3dshape/configs/hunyuandit-mini-overfitting-flowmatching-dinol518-bf16-lr1e4-4096.yaml
index bc3ec2a..193c5ae 100644
--- a/hy3dshape/configs/hunyuandit-mini-overfitting-flowmatching-dinol518-bf16-lr1e4-4096.yaml
+++ b/hy3dshape/configs/hunyuandit-mini-overfitting-flowmatching-dinol518-bf16-lr1e4-4096.yaml
@@ -33,6 +33,22 @@ dataset:
     std: &std [0.5, 0.5, 0.5]
 
     #! Point cloud sampling
+    # Important: The total number of points (pc_size + pc_sharpedge_size) is coupled with the
+    # architecture of the pretrained ShapeVAE in `first_stage_config`. You must ensure that
+    # the total number of points provided by the dataset matches the model's expectation.
+    #
+    # For the default pretrained ShapeVAE (`tencent/Hunyuan3D-2.1` VAE), the model
+    # implicitly expects a total of 81920 points.
+    #
+    # Correct configuration (for the default model):
+    #   pc_size: 81920
+    #   pc_sharpedge_size: 0
+    #
+    # Incorrect configuration that will cause a `split_with_sizes` RuntimeError:
+    #   pc_size: 10240
+    #   pc_sharpedge_size: 10240
+    #
+    # If you need to use a different number of points, you must retrain the ShapeVAE model first.
     pc_size: &pc_size 81920
     pc_sharpedge_size: &pc_sharpedge_size 0
     sharpedge_label: &sharpedge_label true