Some Observations on Generative Art

January 09, 2022

TL;DR: A month into generative art. Framework of generative art (visual complexity, diversity and coherence; organic v.s. computer-like; process v.s. content).

How It Started

I started playing with generative art in early November after going to an event by @tarunchitra and @ruthienachmany at NYC Salon. I loved the talk by @emilyxxie and her artwork so much that I started collecting her work and started some of my own generative art experiments.

Around early December I stumbled upon fxhash. I followed a bunch of generative artists on twitter and started seeing cool geometric and minimal art on my timeline on this platform. Intrigued, I decided to do a quick experiment.

Around the same time Tyler Hobbs was minting “Incomplete Control” in NYC. I read a few of his essays, in particular, Flow Fields and was very inspired. On a gloomy Saturday, I sat in front of my laptop and coded up a flow field, which turned into my genesis collection on fxhash, #3289, Infinite Blue.

How It’s Going

My first collection minted out in a few seconds, which was a huge surprise to me. Then I got hooked, experimenting and iterating on small projects. So far I have done six collections and collected 18 works by some really talented people.

It’s maybe less about NFTs for me and more about creativity. I avoided NFT for a long time because the aesthetics and culture of things like CryptoPunks and BAYC don’t really resonate with me. And I have my own doubts about “web3” that is pretty much in line with what Moxie wrote about: Moxie Marlinspike >> Blog >> My first impressions of web3. Of course he wrote it much better than I can so I won’t expand on that.

But there is some kind of energy in the generative art community, especially on fxhash that’s quite exciting. Similar to the sentiment expressed by Moxie: “but I also understand why nerds like me are excited to build for it. It is, at the very least, something new on the nerd level – and that creates a space for creativity/exploration that is somewhat reminiscent of early internet days.”

And the abstract and minimal aesthetics, the intersection of art and math, the community culture of the generative art is just gold. I am also very humbled by all the talented people entering the space, and am constantly learning and drawing inspirations from them.

The fxhash developer community is super wholesome. Great founder and mods. Tools like fxhash++, the open source APIs, etc. It’s been really fun to develop and contribute.

Some Tentative Framework on Generative Art

I am no art major but as a former academic I have the urge to organize things into frameworks. So this is a tentative framework on my own thoughts on generative art and some open questions. I hope to hear your thoughts on how you conceptualize the space as well. This builds on Tyler Hobbs’ The Rise of Long-Form Generative Art and maybe some scholars on Vera Molnár could point me to prior work too.

Also, this is more focused on the art analysis rather than any price analysis, for which there are great resources such as Kaloh.eth/tez.

Visual Complexity, Diversity, and Coherence

One of the criteria for art that seems to be valued in the community is visual complexity. zancan’s Garden, Monoliths collection is a great example. There is a sense of awe when there are many details in the artwork.


Since the work is usually launched in a collection, people tend to value collections that are visually diverse but coherent at the same time. If the algorithm always generates very similar work, then the difference between each edition is limited and quickly becomes not as interesting. In a way, there needs to be enough entropy in the algorithm generating the artwork to contain more information and value. At the same time, the entire collection must have a coherent theme for them to still be a collection organically. Designing the balance between diversity and cohesion is something I am exploring right now. Maybe I can even write a meta algorithm to evaluate my algorithm generating the art to optimize for a trade-off point between complexity, diversity, and cohesion.

Very Organic or Very Computer-Like

Another feature the community seems to value is the feeling of not being computer generated. It’s kind of ironic, but works by Yazid shows how elegant and organic simple lines, shapes, and colors can be.


In order for the work to feel organic, curves are also very popular – many flow field works in the space. Some most famous collections include: Fidenza by Tyler Hobbs, Wavelength by Kaleb, and many other variations.

Some other work on the platform may instead go to the opposite extreme, which is very computer-like (cyberpunk almost). For example, De/Frag by toxi.


It seems like that the styles in the middle, too obviously computer generated, are not as valued.

Process & The Technical Aspect

Because I come from more of a technical background, the technical aspects and the process of the generative art of great interest to me. When I look at a new collection, I usually inspect the source code and glance at the quality of the code before making a collecting decision.

A “trick” for making the work look more organic is to embed assets or hard code things into the code. For example, one can embed a few jpegs as texture source input, and sample from the images to generate more organic textures rather than synthesizing the noise from scratch.

For example, the beautiful Chaos Research collection by IOivm has two jpg assets embedded.


The Obscure Cameras collection by Orr Kislev has a folded paper as an asset for the background but the rest was purely generated.




For example, in my own work, Inwardness, I did a lot of computation offline in a python notebook and hardcoded the generated palettes in a giant array.

let styles = [
    ['klint_0', 2, 195.33, 163.19, 159.46, 0.0, 0.74],
    ['klint_0', 2, 97.6, 62.41, 63.04, 0.74, 1.0],
    ['klint_0', 3, 187.91, 150.28, 147.29, 0.0, 0.62],
    ['klint_0', 3, 96.44, 60.92, 61.45, 0.62, 0.88],
    ['klint_0', 3, 228.94, 224.9, 217.98, 0.88, 1.0],
    ['klint_0', 4, 204.61, 162.12, 151.62, 0.0, 0.44],
    ['klint_0', 4, 140.42, 115.43, 129.63, 0.44, 0.7],
    ['klint_0', 4, 88.58, 48.11, 42.79, 0.7, 0.88],
    ['klint_0', 4, 229.93, 227.18, 220.79, 0.88, 1.0],

There is a very fine line between “cheating” v.s. being clever. And it’s not an aspect of generative art that people talk about much so far because of the technical barrier. But I think it’s super interesting as the space becomes more and more competitive.

Before, the fxhash projects have a size limit of 15MB. So it’s not feasible to embed many jpeg assets. As the size limit increases, there is more and more space to play with the dynamics of generating on the fly (limited to browser compute and rendering time constraints) v.s. directly importing assets and compressing data.

It will be really cool to see the boundaries people can push technically as well as artistically. There are some WebGL experts in the community that are mind blowingly good. Some other people with oil painting and sketching backgrounds. I think in my own work I tend to bring a data and algorithm perspective before the visual component. And I view myself more as a conceptual artist than a visual artist. So I look forward to pushing the boundaries of generative art, using perhaps network/graph data, tensorflow in the next few projects to provide commentary on our networked age.

Stay tuned!

Profile picture

research engineer by day. generative artist by night.
twitter | fxhash