Figure 2.
The diagram shows a diffusion model block. The input passes through 1 by 1 convolution and Re L U, then enters a residual stack. A diffusion step embedding is processed by a fully connected layer F C with swish activation and another F C, then broadcast over length and added to the main path. A dilated convolution is applied, followed by addition and gated activation using tanh and sigma functions multiplied element wise. Two 1 by 1 convolution layers generate residual and skip outputs. Skip connections are summed and passed through 1 by 1 convolution, Re L U, and a final 1 by 1 convolution to produce the output.The network architecture of unconditional DiffWave