Sample 100 latent codes from normal distribution and input them into the trained decoder:
Sample 100 latent codes from normal distribution and select the label of which image you want to generate, and then concat them to input into trained decoder:
The expected label is 1:
The expected label is 9:
If you adjust the hyper-parameters carefully (such as the structure of the neural network, the dim of latent or the times of train), the results may be better.