Creative Portraiture: Exploring Creative Adversarial Networks and Conditional Creative Adversarial Networks

TL;DR

The paper addresses the challenge that standard GANs may simply replicate training data and explores creative adversarial networks (CAN) and their conditional variant (CCAN) to generate novel portraits. By training on WikiArt portrait subsets at a reduced resolution, the authors show CAN can yield more diverse, expressive outputs than a DCGAN baseline, while CCAN demonstrates style-conditioned but still creative outputs. The study provides a concrete, scalable framework, including training procedures, qualitative assessments, and a public code repository, to advance computational creativity in visual art. It also highlights the need for larger, more capable models to achieve richer detail and raises questions about genuine machine creativity versus learned stylistic reproduction.

Abstract

Convolutional neural networks (CNNs) have been combined with generative adversarial networks (GANs) to create deep convolutional generative adversarial networks (DCGANs) with great success. DCGANs have been used for generating images and videos from creative domains such as fashion design and painting. A common critique of the use of DCGANs in creative applications is that they are limited in their ability to generate creative products because the generator simply learns to copy the training distribution. We explore an extension of DCGANs, creative adversarial networks (CANs). Using CANs, we generate novel, creative portraits, using the WikiArt dataset to train the network. Moreover, we introduce our extension of CANs, conditional creative adversarial networks (CCANs), and demonstrate their potential to generate creative portraits conditioned on a style label. We argue that generating products that are conditioned, or inspired, on a style label closely emulates real creative processes in which humans produce imaginative work that is still rooted in previous styles.

Figures (5)

• Figure 1: Algorithm 1. showing how the CAN is trained. Note that the CCAN is trained in a very similar manner, with the only addition being that style labels are provided to the discriminator and generator.
• Figure 2: Average Generator and Discriminator losses over 120 epochs. Training was stable for all three epochs and the Discriminator and Generator loss did not diverge over 120 epochs
• Figure 3: Selected output from DCGAN. Note the impressive qualities of the output, including the positioning of the central figure, the clothing details. Notice the limited and crude facial expressions.
• Figure 4: Selected output from CAN. Note the diverse style of the output when compared to the DCGAN
• Figure 5: Selected output from CCCAN. Each row is the resultant output of a different style vector. Row 1: Realism, Row 2: Rococo, Row 3: Romanticism, Row 4: Baroque. Note that the portraits in each row have similar styles, yet differing details