Short video showing mode collapse on a StyleGAN2 trained with the details of huipils from Guatemala and Mexico. Even using a SOTA GAN such as StyleGAN2, mode collapse is still a possibility if not enough training data is available and/or we overtrain the model. In our case, we believe the former influenced more than the latter, as our model was trained (using a pretrained model fron the original repository) for under 1 Million images before it started to show signs of mode collapse.

свернуть