Set in Stone

Artist Statement

A series of marble faces, generated by AI, as it learns to create and update its bias’ on gender. First it is trained to generate masculine marble faces, fixed, immovable. Then marble feminine faces are added. It learns to change, a transgender neural network, updating its knowledge and its experience as it goes. Then the marble starts to give way, non-conforming self expression, color, joy, emerging as the gender becomes unfixed, non-binary. Through this evolving conceptual artwork I am exploring how my machine creates images representing gender and if it holds onto its trained bias. I am examining what the gender shift looks like and what transgender and non-binary self expression and self aware aesthetics are beyond biological essentialism. The datasets used by the machine are created by me using 3D modeling and echo the idealized masculinity and femininity of classical statuary with a more inclusive bent. I am purposely training bias into my machine and then attempting to unbias it and see how the artwork evolves as the machine learns that there are multiple genders and gender expressions.

The works you see on the wall and the screen are moments in time in the training of the Deep Convolutional Generative Adversarial Network as I manipulate the dataset it relies on to challenge its bias. As it learns it generates sample images to show its progress and understanding. The samples are entirely represented in the video to represent the full training of the neural network to date. The works on the wall consist of the sample rendered into a mosaic of a face. They are chosen by the application. This is part of my ongoing research project and is an evolving conceptual piece.

Dataset Development

My work is classically inspired in aesthetics to create a resonance with history and historical images. There is a satisfaction to tying the latest in technology with the depth of historical artwork and the anonymizing sense of marble. I desire to create an imagined history for gender minorities, bringing visibility to those largely unrepresented in art history. This involves crafting a historical context for a modern mode of artistic creation.

I used the following multistep method to create the dataset:

1. Mold 3D models of faces or bodies in Daz3D.
2. Create a camera attached to a null object in the 3D workspace that allows stable rotation.
3. Create a large animation timeline of 1000 frames.
4. Create a keyframe at each 100 frames to morph the 3D model into another form.
5. Create an animation sequence where the camera moves around, up, and down to capture as many angles as possible per 100 keyframes.
6. Render the output as images.
7. Flip the images in a bulk image editor to increase their number and present a different view for the GAN.

DCGAN Version

In my case I trained these as single works rather than a grid, the gridded training appears to require more epochs per train as the dataset is small. I am using 1000 images per gender. I created the images in the dataset using 3D modeling and renders to gain different angles and a variety of physical features as the figure morphs between different sculpts and appearances. The marble texture recalls classical sculpture, but also the rigidity of stone of being cisgender and of the opinions of some people about the supposed immutability of gender. The addition of color symbolizes the breaking free of gender constraints, of leaving the stone behind and gaining warmth and color. As the series progresses I anticipate adding further transgender aesthetics in full color, having completely shed the restrictions of stone to realize your self. 

Installation Images

These works are work in development, first shown at Future Histories as a mural and a video work showing the training process. This work was also a finalist in the Midsumma and Australia Post Art Prize

The video shows the full sequence of the DCGAN training and the disruption of data:

ProGAN Version

The next neural network that I used to build on this style of work was ProGAN. With a progressive growing GAN, each sample builds on the one before it, so that instead of generating thousands of individual images, it works with a grid of images and builds on each face in the grid, dividing the pixels until the image is more and more resolved. This process allowed for works which were more suited to a smooth video transition. While the processing is more intensive, each video shows the transition with little to no pixel disruption. This aligns well with the projection work I developed for the Gertrude Street Projection Festival. I took extra steps, dividing each grid into separate parts and carefully disrupting the training. If live manipulation in the DCGAN training was labeled disruption training, this form could be more accurately termed subversion training, as the new data is stealthily renamed to subvert and replace the existing data.

Judith Butler discusses subversion in her book Gender Trouble where she explores making gender trouble through subverting and displacing masculine presences. 

I used ProGAN to create a looping projection based work at Gertrude Street Projection Festival. and Epoch Loop

The images and video below were taken by Bernie Phelan for GSPF 2021

This video shows the full looping project installed for GSPF.