HomeProjectsLearnOur Beliefs
Cover

Flower Construction

GenerativeImageArt

There has been a lot of discussion about whether or not art generated by an AI model is considered art, specifically due to the fact that these models do not specifically “understand” what they are creating; unlike an artist, they do not have artistic intent. In an attempt to understand how to merge AI and art, we partnered with a local experimental artist and researcher.

Upon brainstorming, we decided that we wanted to create flowers, as they show thriving life; however, we wanted our artwork to be abstract. To create this organic feel, we used an unstable DCGAN. While there are options out there that would create more realistic looking flowers, we were not looking for realism, we were looking for abstraction.

The image is a graph titled “Generator and Discriminator Loss During Training”. The x-axis is labeled “iterations” ranging from 0 to 300000, and the y-axis is labeled “loss” ranging from 0 to 60. There are two lines representing the loss values of the Generator (G) and the Discriminator (D). The Generator’s line, depicted in blue, fluctuates between approximately 0 to over 50 in loss value. The Discriminator’s line, depicted in orange, also fluctuates but remains mostly below 20 in loss value. There are noticeable spikes in the Generator’s loss values at various points along the x-axis.

After training for 200 epochs on the Indian Subcontinent Flowers dataset, we were able to synthesize a result that we wanted: realistic upon first glance, but obviously fake upon closer inspection.

A grid of 256 synthesized flowers that look real at first glance, but are fake upon closer inspection.

Training details:

Batch Size: 128

Learning Rate: 2e-4

GPU: Pny Quadro P2200 Nvidia 5gb GDDR5X Graphics Card VCQP2200-PB

GPU Hours: 15

This art piece was shown and won an award at a live exhibition for a university-hosted art show.