auto_NS

auto_NS is a series of interconnected Stable Diffusion models trained on my archive of work that can be used to produce new iterations of videos and still images. Though originally created to fill in the social media gaps caused by infrequent posting, the models also serve as a quick prototyping tool when initiating a new project, they offer me a way to share ideas quickly with minimum effort, and they allow me time and space to do more of the things I love. The models will evolve as my practice evolves.

Below is an essay about the project.

I make work slowly. I like the slowness, it feels like a luxury. There are long periods of time when I don't make art, and others when I don't do much else. I don't talk much about it. I prefer to allow the viewer to decide what the work means to them, if anything. My way of doing things is mostly incompatible with industry expectations for how an artist should market themselves online, but I still choose to make this my living. The modern internet is no place for quiet artists.

For many years I've used automation and generative techniques in my practice. Before readily-available LLMs, I used random text generation to create narratives and meaning for work when there was none. I play with code to create random textures/forms/starting points. I pull external data to define a set of rules/limits to structure what I'm making. I love the unpredictability of random numbers. I love how data can be hidden within art that is more aesthetically pleasing than typical data visualisation. I love creating within finite limitations. It's a place I come back to when I'm stuck. It's a perennially surprising way to make work.

I had the idea to automate the work that I shared on social media in 2017. I know for a lot of people that sharing daily work/process/BTS is a seamless behaviour, but I was annoyed at having to spend time on something that was draining my spirit. I longed for a tool that could follow the steps I take to make work that would ultimately create something that looked like I could have done it, but without my intervention. I wanted to match the emptiness I feel on social media with my own empty work.

That lead me down a path of coding a Python app that would randomly take assets from a folder that contains 1000s of images/videos/3d objects I had used in my work over the years, combine them in different ways, and output an image or a video loop accompanied by a mundane generated CLIP caption. The results weren't amazing, and I didn't have much time to spend on it because of eternal work etc. I had plans to expand on the ways the layers were combined to more closely mirror the way I make things move, but I was also reaching the upper limit of my coding knowledge and so I hit a dead end.

A few years passed. Though I would still daydream of the days I wouldn't have to think about what to post on Instagram, I started to give up on the idea: until I discovered latent diffusion models.

As I understand it, latent diffusion models take training data, create a simplified representation of it categorised by its defining features [latent representations], create new latent representations that look like a mix of the original ones, destroy the data by adding noise until there is nothing left but noise, then they remove the noise [denoising autoencoder] to create new instances that resemble the training data [synthetic images].

So instead of using existing models trained on the billions of images scraped from the internet, I trained my own model on my own work. I took the same 1000s of assets from my failed Python app and around 24hrs later the first version of auto_NS was complete.

Out of ≈50 generations I will pick one. It doesn't feel like my own work but it looks like it. I feel like the curator, not the artist. It feels freeing. I don't have to make any constructive creative decisions, I just have to decide 'yes, keep' or 'no, delete'. If no one else likes it, it doesn't feel personal. It's filler, albeit comfortingly beautiful filler. And it takes max. 30 minutes of my day.

The accompanying generated captions are because, no matter how trite and forced, all things shared related to your art practice have to have an explanation/context, even though most things around us don't have an explanation/context, just our internal interpretation of those things. I especially love how painfully boring CLIP captions can be, I feel it enhances the nonsense of it all.

There is a large part of me that doesn't want to contribute anything new to the already overwhelmingly infinite amount of stuff to look at. There is enough of it already. I am tired. My eyes are tired.

auto_NS is the response to that feeling. A way to free up my time again: for slow work, for motherhood, and for being IRL, mostly.

Natalia Stuyk
[26-02-2024]