Data Generation and Latent Space Based Feature Transfer Using ED-VAEGAN, an Improved Encoder and Decoder Loss VAEGAN Network
Jiatong Li · Atlantis Highlights in Intelligent Systems/Atlantis highlights in intelligent systems · 2023
To combine the advantages of VAEs and GANs to generate both diverse and high-quality samples, this paper proposes ED-VAEGAN which improves encoder and decoder loss of traditional feature-wise VAEGAN [4].More precisely, a reconstruction score term is added to encoder loss function, which accelerates the training of the whole model.The decoder loss was similar to traditional definition, but discarded an irrelevant term to decoder.This paper applied this new model to face datasets and compares the generations with other models when the models are fully trained and when trained for the same iterations.And the latent space expedition was done by first encode the images and then do the latent code walk between two images.As a result, ED-VAEGAN outperformed traditional VAEGAN on training speed, and its latent space expedition result indicates better continuity comparing to other pixel-wise models.In the end, this paper applied simple data augmentation method to solve the brightness problem that happened when training iterations increase.