Memleak.nude: a digital installation created through the nakedness of metal.

Thesis and Methodology:

Humanity has switched intimacy and its entire realness into digital life, whereas public life has became a source of self-censorship. Our silicon collects it all, unfiltered, unlabeled and untethered by human boundaries. The unfilteredness of data enticed me. Even though computers in the top levels have context about the data, like file headers, everything is clear and unfiltered binary when looked upon. Therefore for 12 days, 20 minutes and 43 seconds before creating the dataset, the artist lived through their computer, never closing anything. allocated. garbage uncollected. then, with aid from ChatGpt learned to dump all the memory into a raw data file, and a python script to convert the raw data to images.

The script generated 47365 images derived from my digital memory. This was a nude self portrait of my naked metal, my body without organs, the most sensitive data leak for a digital body.

The artist then curated 102 images out of this massive data to train this embedding. both as a way to keep their sensitive data safe and to be anonymously digitally naked. The embedding was overtrained with a small learning rate to capture as many details from visualized raw data, 10000 steps and 1e-6. every third, the model was changed, sd 1.4, analog diffusion and retrodiffusion in the final pass.

the model and the embeddings are cc0, any outputs that come out can be commercialized by the output's creator, the end user, with no reference needed to me, the author. have fun, do anything you like with it. i can't impose restrictions on beauty.

here are some of the raw outputs:

Intended uses & limitations

create compositions out of an artist's most sensitive data leak.

How to use

copy the contents to your inversion embedding folders for your local stable diffusion.

Limitations and bias

the embedding works best at creating init images and be used in AND statements with a weight. i.e. a cat AND nakedMetal::0.ab . Personally found 0.17-0.4 the best range but you are invited to experiment if using.

Training data

the training data consists of 102 images of the artist's personal sensitive raw data. here are a few examples: