Label arithmetic forms intermediate conditioning labels by weighted averaging two labels. In a conditional GAN, these labels guide a generator between attributes, such as faces with and without eyeglasses, so the attribute changes gradually. section 5.0 section 5.6
It differs from vector arithmetic: label arithmetic changes the conditioning label, while vector arithmetic averages latent noise vectors. Combining both controls separate characteristics at once, such as eyeglasses and male-to-female facial features. section 5.0 section 5.5.3
Fading eyeglasses with interpolated labels
Create labels_g for eyeglasses and labels_ng for no eyeglasses. For weights 0, 0.25, 0.5, 0.75, and 1, compute label = w*labels_ng + (1-w)*labels_g. Concatenate each reshaped label with the same latent vector, pass it to the trained generator, and plot the five outputs. The glasses gradually fade as the weight moves toward the no-eyeglasses label. section 5.5.1
The same idea scales to a grid: interpolate the label across columns and the latent vector across rows. A 6×6 grid can therefore show eyeglasses fading horizontally while facial features change vertically. section 5.5.3
