Does the number of biases in a Convolutional layer scale with the number of images?

Question

I am convolving 32 grey scale images of size 28 x28 with 16 filters of size 5x5.

Which of the following is the correct way to add biases to the convolution operation output?

1. add 1 scalar value bias for each of the 16 filters. (Each of the 32 images share the same weights and 16 biases)

2. For each of the 32 images, add 1 scalar value bias for each of the 16 filters. Each image has its own weights and 16 biases?

3. something else- Please explain