Fig. 3: Training of different generative models for synthetic cell dataset creation.
From: Deep generative modeling of annotated bacterial biofilm images
a VAE implementation with convolutional layers yields more realistic images. b Learning loss curves for the convolutional VAE. c Comparison of images generated by WGAN with linear and convolutional layers. d Training losses of the discriminator and generator for WGAN with convolutional layers. e Example of images generated using the diffusion model and training loss curve. f Comparison of the FID metric for the implemented generative models: WGAN achieved much better results than the other models.
