Different keras versions generate different architectures of the same model.. WHY?

Question

Please i am facing problem in building the same model using different Keras versions.
Note that the code was written in Keras 1.x and I am working on reproducing it using Keras 2.x.

the code in Keras 1.x :

 a = Input(shape=(1, 512, 512),name='a')            #a is the original groundtruth
 b = Input(shape=(3, 512, 512),name='b')            #b is the original image
# A -> B'
bp = atob(a)       # this is to generate fake image (B') from ground truth (A) using atob network
# Discriminator receives the pair of images
d_in = merge([a, bp], mode='concat', concat_axis=1)
pix2pix = Model([a, b],  discriminator(d_in), name='d')

the output of Keras 1.x is

while in Keras 2.x using the same code

a = Input(shape=(1, 512, 512),name='a')
b = Input(shape=(2, 512, 512),name='b')
# A -> B'
bp = atob(a)
# Discriminator receives the pair of images
d_in = concatenate([a, bp], axis=1,name='concat')
pix2pix = Model([a, b], discriminator(d_in), name='d')

the output is

the Question: why there is an additional layer named InputLayer (b) appeared in the output of Keras 2.x and not appeared in the output of keras 1.x ?

Note: the full code of keras 1.x can be found on Costa_github_line388

Thanks in advance.