Generated patterns

Programming Patterns: AI Generated Patterns

This work explores the current capabilities of machine learning technology for generating new patterns based on input data consisting of 18th-century fabric archives. The main focus is on the principle of GAN architecture function, dataset generation issues, and the training process. The work highlights the network of human decisions and steps behind the seemingly automatic process.

① Patterns / Online Archive
In 2021 was involved in a research project called Beauty Patterns, led by Assoc. Prof. František Svoboda from Masaryk University. The project aimed to digitize an extensive archive of fabrics from the Brno dioceses dating back to the 18th century. My task was to create a digital archive to display these patterns. While working on this, I found myself constantly thinking about the patterns and starte to explore how to make this valuable historical material more accessible and engaging for modern audiences.

② Jacquard Loom and Punch Cards
I became more interested in the origins of the patterns themselves. This is the Jacquard machine that was used to create the patterns. With its punched cards, it simplified a time-consuming an labor-intensive process. It is often considered the forerunner of modern computing because it interchangeable punched cards inspired the design of the first computers developed by Charle Babbage and Ada Lovelace.
I was fascinated by the notion of machines tackling complex tasks, prompting me to explore additional technologies for uncovering new potential for the patterns archive.

③ StyleGAN 2
I decided to utilize NVIDIA's StyleGAN 2 model, which generates synthetic replicas based on inpu data. It has garnered significant media attention in the past for its project focused on creating fictitious (fake) human faces. 
④ Curating Dataset
Initially, the first results were just pure glitches, so I needed to dig deeper into the problem. Quickly, I discovered that the reason for the failure was the dataset. For any type of machin learning, the dataset is essential. This refers to the set of data that the neural network processes to produce new results.

The entire dataset had to be curated; the original cuts had to be handpicked or post-productio edited, and the overall color scheme had to be monitored.

I created the dataset in the middle of ongoing research so it was also not completely reported and therefore I had to go through about 6000 photos from which I was able to extract 1105 patterns. 

⑤ Synthetic Patterns Generation
I began a new training session using a carefully selected dataset, and it produced excellent results. New synthetic patterns were consistently generated, but I found myself feeling overwhelmed by the amount of data being generated. Despite this, I continued to be amazed by the new patterns. I made interesting observations about the color choices and the consistent flexibility of the patterns in th generated images.
The capabilities of networks like StyleGAN 2, which I've used in my work, are incredibly impressive. However, from what I've verified, it's still an intelligent recyclation of a dataset rather than true artificial intelligence. It appears that we are still quite far from dystopian visions in which machine do everything for us, though we can already see small indications.

⑥ Results
3 months of local training / generated 94 of intermediate steps
/ 1 steps = over 4 billion patterns / total xxxxxxxxxxxxx billion sythetic patterns 
⑦ Videos
Technology allowed me to create videos within the latent space, bringing movement to the static archive and reviving the original baroque flower motifs.

⑧ Fashion collection Creating a dataset from something handmade and physical made me think about how to transfer generated data back to the physical state. So, I teamed up with fashion designer Aleš Hnátek to create a real fashion collection from the patterns I generated.