SNcGAN - Generate Conditional Images

Spectral Norm + Conditional GAN

Adeel Mufti, Biagio Antonelli, Julius Monello
Link to paper:
GitHub repo:

This is a demonstration of a Generative Adversarial Network with Spectral Normalization[1] that has been conditionally trained[2] on images of oil paintings of faces of people, extracted using OpenCV, from the PainterByNumbers[3] and BAM[4] datasets, with conditioning labels created using Microsoft Face API[5]. Additionally, a SNcGAN trained on the faces of celebrities from CelebA[6] dataset is demonstrated.

The form below generates images from a live TensorFlow model, using the conditional labels chosen.

[1] Miyato, Takeru, et al. "Spectral normalization for generative adversarial networks." arXiv preprint arXiv:1802.05957 (2018).
[2] Mirza, Mehdi, and Simon Osindero. "Conditional generative adversarial nets." arXiv preprint arXiv:1411.1784 (2014).
[3] Duck, Small Yellow. "Painter by numbers," Kaggle (2016).
[4] Wilber, Michael J., et al. "BAM! the behance artistic media dataset for recognition beyond photography." Proc. ICCV. Vol. 1. No. 2. (2017).
[5] Microsoft Face API
[6] Liu, Ziwei, et al. "Deep Learning Face Attributes in the Wild" ICCV (2015).

This project won 2nd place out of 124 projects in a competition hosted by IBM.
Our conditional GAN architecture

Celebrity Faces

y-vector (conditioning label)
Generated Samples

Face Paintings

y-vector (conditioning label)
Generated Samples

Real samples from CelebA training dataset
Real samples from face paintings training dataset