It seems that the transformation of color noise does not look like color noise at each point, but simply makes a random change over 3 channels common to all points
https://nbviewer.jupyter.org/github/goodok/fastai_sparse/blob/master/notebooks/transforms/transforms.ipynb
x_colors = pc.colors
amplitude = 0.1
x_colors_b = x_colors +
np.random.randn(3).astype(np.float32) * amplitude
x_colors - x_colors_b
returns differ between before and after:
array([[0.03873268, 0.03023028, 0.1048553 ],
[0.03873268, 0.03023027, 0.1048553 ],
[0.03873268, 0.03023028, 0.1048553 ],
...,
[0.03873268, 0.03023028, 0.1048553 ],
[0.03873265, 0.03023028, 0.1048553 ],
[0.03873265, 0.03023028, 0.1048553 ]], dtype=float32)