Generative AI models been around for a few years now, progressing in strides to ceate the most picture-perfect images possible. But where GANs aim to get the most minute details right, I wanted to capture the vague essence of what an insect is. What makes it so recognisable, outside of the scientific definition? How would you imagine an insect from memory?

A quick Google search returns hundreds of tutorials on how to create your own generative model. An Autoencoder seemed like the simplest, but also most project-accurate option: the encoder takes the images from the dataset, shrinks it, and passes that to the decoder that will try to "re-imagine" what the image was like.

A bit of time spent on Kaggle gave me more photographs of bugs than I could have hoped for, and I went on to build a decent dataset. Beetles being basically evolution's favourites, they make up a little over half of my dataset. Natural History Museum drawers have a huge variety of bugs, but they all look a sad shade of duty brown-grey, so I injected a handful of ultra-vibrant butterflies and aphids into the mix. Datset biases account for a lot of issues related to AI, but in this case it helps support my project all while exposing the impact human selction has on the results.

Then came the training, figuring out how many epochs give the best result - not too vague, not too close - but also how to upscale the images and give them a soft, oniric quality.