wuhuikai/GP-GAN

Implements a Wasserstein GAN-based approach with both supervised and unsupervised variants to seamlessly blend source and destination images given a mask input. The architecture uses an encoder-generator-discriminator framework optimized for high-resolution compositing, with pretrained models available for immediate inference. Built on Chainer 6.3.0, it supports training on the Transient Attributes Dataset or unsupervised learning on landscape imagery from the Places dataset.

477 stars. No commits in the last 6 months.