Vector arithmetic interpolates between generated attributes by taking weighted averages of noise vectors in latent space. In a conditional GAN, the generator turns these interpolated vectors into images that transition between two instances, such as male and female faces. section 5.0 section 5.5.2
It differs from label arithmetic: vector arithmetic changes the selected noise vectors, while label arithmetic changes conditioning labels. Combining both lets a generator vary independent characteristics at once, such as facial gender and eyeglasses. section 5.5.3 section 5.6
Interpolating between two faces
Start with z_male_ng and z_female_ng, then choose weights 0, 0.25, 0.5, 0.75, and 1. For each weight w, calculate z = w*z_female_ng + (1-w)*z_male_ng. Reshape z to (1, z_dim, 1, 1), concatenate it with a label reshaped to (1, 2, 1, 1), and pass the result to the trained generator. The resulting five images progress from male features toward female features. Replacing labels_ng with labels_g repeats the interpolation with eyeglasses. section 5.5.2
Varying two characteristics
Use six interpolation values for the noise vector and six for the label to create a 6×6 grid. Moving down the columns changes the face gradually from male to female; moving across the rows makes eyeglasses gradually fade. section 5.5.3
