The direct inspiration for the project came last year when I was visiting my parents in Lithuania, staying at their house outside Vilnius, with a view from my window over the forest that dressed beautifully in fog during early dusk hours. In many of my projects, I explore the theme of “home” and what it means emotionally. And I must say that I subconsciously associate pine trees and fir trees with the notion of familiarity and calm. There is also a part of me that spent summers in the deep Lithuanian forests at underground psytrance and techno music festivals. The colorful hand-made decor of these festivals fitted so masterfully with nature as if it was an illustration of godly forces and souls inhabiting these places.
Mythic Latent Glitches are a continuation of the explorations that I started with Forest Gods, where I juxtapose the mystery associated with forests in folklore with the mystery of artificial intelligence networks that can create real-looking photographs of forests and invent motion transformations between still images. Pagan gods and fairies are deeply rooted in Lithuanian folklore and several mythological creatures were said to be living in deep forests spanning the country. As the Baltics were one of the last pagan nations in Europe to Christianize, mythology from the area is a lot more recent than in other areas. Maybe that is also the reason why many of the Pagan rituals are still present in our culture, even if on symbolic and cultural levels (I welcomed the year 2022 by throwing amber dust into a giant campfire, an old revived tradition). 
While I was researching the Lithuanian pagan gods in the folklore, I found much conflicting information, as it often happens with scarce documentation and stories mostly shared from mouth to mouth.

Pagan rituals. Source: medievalists.net

From the myriad of gods and goddesses and nature spirits, I selected a few that I felt could inhibit the contemporary generative forest forms:
Žiburinis - a phosphorescent forest spirit
Žemyna - the goddess of Earth, is usually regarded as a mother goddess and one of the chief Lithuanian gods. She nourishes all life on earth, human, plant, and animal. 
Medeina - the goddess of hunting 
Auštra - the goddess of the wind, that stands at the gates to paradise.
Mėnuo - Moon, the god’s son.
Vakarinė -  the goddess of the Evening Star.​​​​​​​
Humans have been always fascinated with metaphysical forces. While old beliefs created anthropomorphic creatures to represent inexplainable stochastic processes, we nowadays associate much of the same mystery, fascination, and fear with technology and artificial intelligence.  
The choice of the “glitch” technique symbolizes the confusion caused by not being able to comprehend and the malfunction of our perceptive processes.

In my creative process, the relationship between my inputs and the technology usually progresses in a multi-stage manner.  In this project, my goal was to combine the human and generative inputs, in a counterintuitive manner, by allowing AI to design the organic elements and use the human hand to incorporate artificial glitches. What seemingly emerges as natural forms, was imagined solely by GAN, seeing the world through the limited examples through the camera's lenses.
First, as in all of my projects, I started with photography. Collecting training data, shooting in the wild, taking lonely walks. I usually perform a mix of images from my archives, memories captured on digital and analog films, and taking photographs with a specific project in mind. When training deep learning models, it is essential to capture a stylistic coherence in the training data and not include too many entangled concepts. Hence shooting for an AI project would already be different from regular photography. For Mythic Latent Glitches, training images span months and years of different experiences.
Below are a few of the original archive photographs that were used to prepare the training dataset. These are my memories of walks in forests surrounding Vilnius, teenage trips to the seaside, and weekends surrounded by music and friends in the wild. To reduce the data ambiguity and complexity, these pictures were afterward intensely processed and cropped.
Carefully curating the input selection that would go into the model is the fundamental building block of my process and it allows me to inject bits of my personality and my style into the machine learning process. 
Afterward, I give full control to the machine to interpret this vast set of data and learn from it, guided by the optimization algorithm.  I use custom code which is a modified version of StyleGAN2-ADA and which has also been continuously trained and re-trained on different kinds of new datasets. With each new project, I like to take my previous model and make it forget or re-imagine old concepts as new, regardless of how different the captured images look. 

Generated image seed = 161

Generated image seed = 1337

While a trained model can span a technical infinity of possible outputs, I like to explore the outcomes in a meditative way to find those that resonate with me and my vision the most. For Mythic Latent Glitches, I chose twenty-three seed numbers, that served as the canvas for further generative explorations. 
This part of the process of selecting the finite set of outputs is also a crucial element of the human-in-the-loop procedure. While machine creates images, the two of the most crucial pieces of information are provided by human: what to create and when is the work done?
As Harold Cohen said: "I give the machine its identity. It is doing what I have in mind".

Generated image seed = 1267

Generated image seed = 1262

For the next part of the algorithm,  a set of stochastic manipulations are programmed to create alternative narratives through visual destruction. There are a couple of main glitch manipulation methods that are carefully applied in different combinations and with different parameter values. For each of the iterations, between 1 and 4 generative transformations are applied.
Around half of the produced iterations have perspective warping that produces unexpected shape patterns and entangled tree branches. There are 10 different offset levels that produce different types of warping.

Warp threshold = 300

Warp threshold = 1000

Another type of applied transformation is the scanned lines. It has a variable noise parameter that prevents the lines from being perfectly parallel to a bigger or lesser extent. The starting position is also variable and in some cases, it obfuscates up to 90% of the image.

Stroke = 1

Stroke = 5

A huge destructive effect and contextual shift are achieved with the use of color burn, with a random choice of threshold color. 

Color burn at R: 1, G: 23, B: 12

Color burn at R: 68, G: 42, B: 159

Color noise is added together with color burn or in some instances as a standalone transformation. It depends on the background color.

Green noise texture (close up)

Magenta noise texture (close up)

Random tile misplacement reflects on memory fragmentation and disintegration as well as on the building blocks of the perception itself, similarly to my previous project "Study of atom in analog, digital and A.I." 
The number of tiles is selected randomly between 0 and 20 unless the number of all applied transformations after the pipeline is less than one. Then, 50 tiles are scrambled.

Number of tiles = 6

Number of tiles = 50

Finally, TV lines are added, where their appearance depends on the position in the transformation pipeline. 

TV lines with tiles.

TV lines with tiles and color noise.

As a special feature, around 30% of all pieces are animated and allow to create a random set of limited scope transformations live in the browser. Viewer is invited to participate in the creative process by stopping the process at an arbitrary moment and exploring the additional possibilities of the algorithm.


Working with long-form generative art has definitely changed the way I think about the generative art algorithm. It was scary but it was also invigorating to leave the final stage of creation up to the random forces and see the tokens being minted by my dear collectors.
A week before the collection launch, I had the pleasure to present a preview of my work at a solo show in HT gallery in Warsaw. It gave viewers a glimpse into the upcoming works of the algorithm and for many, it was an introduction to the puzzling concept of on-chain output generation, guided by the random process.
The full collection of 333 tokens is now available on (fx)hash.
Exhibition Latent Glitches on view 3-13 July.
//side note
I’m a mixed media artist and I find my voice in combining analog and digital techniques. However, it must also be noted, that Mythic Latent Glitches is fully creatable with code. No pixel in the final outputs has been hand-drawn or captured on camera. The use of base images serves as offline memory optimization and is equivalent to prompting my custom GAN model with a list of pre-defined seed numbers (it is possible to create all outputs online in pure code with a GPU card and ~200 MB of memory). It is a very different technique from the traditional procedural art and will perhaps be recognized accordingly in the future. Nonetheless, I am excited to work in a computer art field that is just emerging, that combines a strict set of rules and algorithms with a more unpredictable optimization output, such as a generative adversarial network.
Back to Top