41.622°N, 71.484°W Camouflage

This project uses an AI image-making model (a latent text-to-image diffusion model) trained on a custom dataset to generate a localized and seasonal camouflage pattern. The dataset is comprised of around 400 of my own photos taken at Davis Memorial Wildlife Refuge in East Greenwich, Rhode Island. The refuge is 40 acres of forest and wetland with a small trail that brushes up against the Hunt river. It's the smallest of fourteen wildlife refuges owned by the Audubon Society of Rhode Island, making it the smallest actively conserved piece of nature in the smallest U.S. state.

AI outputs are often measured by their ability to eliminate anomalies or inconsistencies — glitches — in their output images. In the case of making a camo pattern, these unwanted abstractions become functional. The glitches and repetitions are exactly what's needed to make a pattern that can serve its purpose as you move through different parts of an environment. The method involves using a dataset to provide supplemental training to an already generally capable base model. This flexibility to "learn" the features of an arbitrary dataset is what allows the outputs to be local and seasonal.

With this exploration I hope to offer an alternate view on the use of AI for image-making — that this is a tool affording highly personalized designs that weren't previously possible. This camo is intended to be used for hunting and fishing, activities which demand a deep understanding of an environment, and which at their best represent a non-naive, hands-on stewardship of our natural environment.

The dataset was composed of ~400 images taken at Davis Memorial Wildlife Refuge in late November. This was used to train a Stable Diffusion 1.5 model using the LoRA method (Low Rank Adaptation of Large Language Models). Around 10% of the outputs were functional as camouflage, from which the final pattern was selected for its naturalism and flexibility.