ControlNet Zero-Conv Params

Problem

ControlNet adds trainable zero convolutions whose weights and biases start at zero. Count their parameters: out_channels * in_channels + out_channels.

The signature is zero_conv_init_params(in_channels, out_channels). Return an integer.

Hint

Zero initialization makes the control branch an identity at the start of training.

Test cases

case 1

in: [320,320]

expected: 102720

case 2

in: [1,1]

expected: 2

case 3

in: [64,128]

expected: 8320

case 4

in: [256,256]

expected: 65792

Starter codePython

def zero_conv_init_params(in_channels, out_channels):
    # Your code here
    pass

4 test cases · run your code against them in the editor