diff --git a/assets/img/voxcpm.png b/assets/img/voxcpm.png
new file mode 100644
index 0000000..c1dee57
Binary files /dev/null and b/assets/img/voxcpm.png differ
diff --git a/src/models/voxcpm/audio_vae.rs b/src/models/voxcpm/audio_vae.rs
index 4b68b15..ccaa914 100644
--- a/src/models/voxcpm/audio_vae.rs
+++ b/src/models/voxcpm/audio_vae.rs
@@ -42,6 +42,11 @@ pub struct CausalConvTranspose1d {
     output_padding: usize,
 }
 
+// 元素间：stride-1
+// 两边： k-p-1
+// (h-1)*s -2p+k
+// (h+1)*s
+
 impl CausalConvTranspose1d {
     pub fn new(
         weight: Tensor,
@@ -161,6 +166,7 @@ impl Snake1d {
         Ok(Self { alpha })
     }
 
+    // x + sin(alpha*x)^2 / alpha
     pub fn forward(&self, x: &Tensor) -> Result<Tensor> {
         let dims = x.dims();
         let x = x.reshape((dims[0], dims[1], ()))?;
@@ -192,7 +198,7 @@ impl CausalResidualUnit {
         kernel: usize,
         groups: usize,
     ) -> Result<Self> {
-        let pad = ((7 - 1) * dilation) / 2;
+        let pad = ((kernel - 1) * dilation) / 2;
         let block0 = Snake1d::new(vb.pp("block.0"), dim)?;
         let block1 =
             WNCausalConv1d::new(vb.pp("block.1"), dim, dim, kernel, dilation, pad, groups, 1)?;
diff --git a/src/models/voxcpm/model.rs b/src/models/voxcpm/model.rs
index d3c03a5..2ed367a 100644
--- a/src/models/voxcpm/model.rs
+++ b/src/models/voxcpm/model.rs
@@ -598,12 +598,10 @@ impl VoxCPMModel {
             inference_timesteps,
             cfg_value,
         )?;
-        println!("laten_pred: {}", latent_pred);
         let decode_audio = self
             .audio_vae
             .decode(&latent_pred.to_dtype(DType::F32)?)?
             .squeeze(1)?;
-        println!("decode_audio: {}", decode_audio);
         let decode_audio_len = decode_audio.dim(D::Minus1)? - 640 - 640;
         let decode_audio = decode_audio.narrow(D::Minus1, 640, decode_audio_len)?;
         Ok(decode_audio)