Training a network to generate images - how to build a network?

Question

Python

Hello!
I'm trying to build a network on the tflearn library
Purpose: generating images based on the data fed to it in the learning process.
Images: own (set of pictures).
As I understood from the description of such generators, the network is trained so that the output data matches the input data, and the network topology itself is like an hourglass - many neurons, then convolution, then the narrowest part needed for subsequent generation, then mirror inverted layers.
That is, a picture is given as input, the network somehow compresses it into a set of features, and then generates an image in the next layers. And the learning process comes down to ensuring that the generated image becomes similar to the input.
Then a half with input data is “bited off” from the trained network, and the generator works on the data coming to the narrowest part.
But I can't build a network in such a way that it uses the input data for training.
I would be very grateful for some simple examples. I saw implementations for generating numbers in a similar way, but did not understand how they work (many letters in the code).

Answer 1

You described an autoencoder, this is far from the only architecture suitable for your task, try looking at GAN.
There are a lot of implementations on github.