Image augmentation looks significantly thicker

[jeffDebug]

Hi all,

When I was applying keras (3.4.1) augmentation APIs (RandomRotation, RandomZoom layers), I noticed that the strikes in augmented images are significantly thicker (see screemhots for typical results). My core code is given below. I tried to play with arg "interpolation", too. Did I miss anything when calling these augmentation API?

`
import keras
from keras.layers import RandomRotation, RandomZoom

data_augmentation = keras.Sequential(
[
RandomRotation(factor=(-0.05, 0.05), interpolation="nearest"),
RandomZoom(height_factor=0.1, width_factor=0.1, interpolation="nearest"),
],
name="data_augmentation",
)

x_aug = data_augmentation(x_trn).numpy()
`

where x_trn's shape is (batch, 28, 28, 3), which is the MNIST digit data with all 3 channels identical.