Problem - 3D Array
Questions: Nov 2012, Oct 2013
import numpy as np
a = np.random.random((2, 100, 4))
b = np.random.random((2, 100, 4))
c = np.random.random((2, 100, 4))
Solution - dict uniqueness
For my case _nolegend_ (bli and DSM) would not work, nor would label if i==0.  ecatmur's answer uses get_legend_handles_labels and reduces the legend down with collections.OrderedDict.  Fons demonstrates this is possible without an import.
Inline with these answers, I suggest using dict for unique labels.
# Step-by-step
ax = plt.gca()                      # Get the axes you need
a = ax.get_legend_handles_labels()  # a = [(h1 ... h2) (l1 ... l2)]  non unique
b = {l:h for h,l in zip(*a)}        # b = {l1:h1, l2:h2}             unique
c = [*zip(*b.items())]              # c = [(l1 l2) (h1 h2)]
d = c[::-1]                         # d = [(h1 h2) (l1 l2)]
plt.legend(*d)
Or
plt.legend(*[*zip(*{l:h for h,l in zip(*ax.get_legend_handles_labels())}.items())][::-1])
Maybe less legible and memorable than Matthew Bourque's solution.  Code golf welcome.
Example
import numpy as np
a = np.random.random((2, 100, 4))
b = np.random.random((2, 100, 4))
import matplotlib.pyplot as plt
fig, ax = plt.subplots(1)
ax.plot(*a, 'C0', label='a')
ax.plot(*b, 'C1', label='b')
ax.legend(*[*zip(*{l:h for h,l in zip(*ax.get_legend_handles_labels())}.items())][::-1])
# ax.legend()   # Old,  ^ New
plt.show()