What was the development process of the project? How did the original idea evolve?

Max: During a design theory course while I was studying in Berlin Weissensee, we discussed the use of artificial intelligence in the design process in depth. I found these theoretical considerations incredibly enriching, but I also wanted to try it out in practice. What would happen if I tried to design something with the support of AI?
That was at the beginning of 2019.

I then started experimenting with various algorithms, all aimed at generating images of Moroccan carpets. I was fairly familiar with Moroccan rugs and, thanks to my father, had access to a large archive of rug images. To be honest, the results of these first experiments were mediocre.

The main issue was data. Back then—and still today—it holds true that a large amount of data is crucial for training AI.
Nevertheless, it was an interesting experiment, one that inspired long discussions between Gebhart and me about the parallels between how AI learns and copies compared to humans. We also talked about how this could relate to the history of rug design in Morocco and whether the algorithm’s designs could be brought back into the craft.

Ida: For me, the idea of the exhibition came from a desire not to limit the outcome of my master’s thesis to just a degree and a nice book. In it, I explored patterns and their development in Moroccan carpets, as well as AI-generated derivations of those patterns. I came across the topic of analogue and digital pattern development in the summer of 2022. While talking with friends about finishing university, the topic of “Moroccan rugs”—a major but until then passive presence in my life—emerged as a potential focus.

What always fascinated me most was the abundance of colourful and seemingly chaotic patterns, derived from rug models from urban manufactories. They evolve from piece to piece, and in doing so, the carpets seem to tell their own development story. A story of great geographical scope: the patterns of these rugs don’t just migrate within Morocco, and not only recently, from urban to rural areas and vice versa. In Moroccan rug-making, in addition to the traditions of the Berbers and Arab nomads, there are also Moorish-Andalusian, Ottoman and other Oriental influences, crossing the Mediterranean. This affects both the craft and the knotted patterns and compositions, which are taken up and transform regional design languages.

This analogue evolution of patterns, which fascinates me so much, is therefore nothing “new” or “modern”—but that of an AI is. Still, how does AI pattern development actually work? Could it perhaps be imagined similarly to the analogue migration of motifs in carpets?

These kinds of questions led me to my brother Max, who was exploring a similar topic from a different perspective. Thanks to his background, he was quick to agree to handle the technical side of programming the AI. It then provided both data for my master’s thesis and the images for our joint exhibition MLrug. Our father Gebhart also quickly came on board as a lexical advisor and supporter of the practical realisation.

How did you implement the project technically? What shaped this process in particular?

Max: In my first experiments in early 2019, I played around with different algorithms. They were all based on so-called “generative adversarial networks”, or GANs. At the time, they represented the state of the art. I achieved my best results using “StyleGAN”, a type of GAN developed by Nvidia.
When I wanted to train a new algorithm with Ida for her thesis, I reviewed which algorithms might be suitable. OpenAI had already introduced DALL·E 2, an image generator based on transformers—a fundamentally new and improved technology compared to GANs.

In the end, however, we chose StyleGAN2-ada because it was much more accessible. Transformer-based algorithms require significantly more data and computing power to train. From my earlier experiments, I already knew data would be a key issue, so we specifically opted for StyleGAN2-ada. This is a variant designed for training with fewer images.
Using it, we achieved results of a quality that honestly surprised me. We trained the model on a cloud platform, as we didn’t have the necessary computing resources at home.

Working with AI always raises the question of training data sources. How did you approach this issue?

Max: Most of the images came from Gebhart’s private photo archive, which he had compiled over the years. This was a very important element, as the quality of these images is also very high. Both the quantity and quality of the training data are crucial.
In addition, we used some images from the internet. It was important to us that these images were only used to train an algorithm within a purely scientific, artistic and transformative context.

Ida: This question also touches on how one handles cultural heritage. In Morocco, copying patterns as a source of inspiration and for learning the craft is culturally inherent. This is how techniques, patterns and compositions are passed on and transformed. The notion of “image” rights is perceived very differently. With MLrug, our aim is not to copy, revolutionise or replace Moroccan rug culture. It’s about opening a new window and seeing where it might lead.

What inspired you to take the step from the digital realm back to traditional craft?

Ida: On the one hand, we come from a home full of Moroccan carpets, so the step was obvious. On the other, I didn’t want the results of my master’s thesis to remain purely theoretical, bound in a book. Once I had grasped the content intellectually, I wanted to comprehend it with my whole body—as a handcrafted object that can be touched, that sparks conversations when seen, and in doing so keeps the themes discussed in motion. It represents the content and embodies the fundamental connection between digital technology and textile craftsmanship.
And of course, curiosity played a major role…

Gebhart: Curiosity. About what AI can read and understand—and what it can’t. Even more curiosity about how to translate that back into craft. The current producer landscape in Morocco is rhizomatic—many small businesses working locally and interconnected. I’ve been collaborating with one of these producers in the western Middle Atlas for about seven years, and together we’ve worked to achieve a level of quality that also enables projects as complex as this one.
Last but not least: a special curiosity about collaborating professionally with my own children.

Choosing the right place for the artisanal implementation is surely an important decision. What guided your thinking there?

Gebhart: For a highly faithful reproduction, a production in India or Nepal would have made more sense, as their manufactories have in-house IT departments that can create computer-generated knotting templates, enabling nearly photographic precision. But we opted for a production in Morocco, with all its improvisational imprecisions, because this approach better reflects the character and vitality of the original material—and also ensures that the value creation remains in the country of origin of the design references.

What unexpected moments did you experience when your digital designs were translated into physical objects?

Gebhart: Since I’ve worked with the production team for several years, I was well aware of their capabilities. Still, I was positively surprised by the liveliness of the realisation, which depends on the interplay of many details.

Ida: My approach was driven by curiosity and by a concept that didn’t prioritise aesthetics. I wanted to bring a long, theoretical process to completion and to share it with others.

Traditional craft and digital technology often seem to exist in separate worlds. What connections have you discovered between them?

Ida: These worlds aren’t so far apart. I want to go beyond the obvious parallel that a weaving draft resembles a pixel image. There’s a deep historical link, and for a first glimpse, it’s enough to listen: “(…) the Analytical Engine weaves algebraical patterns just as the Jacquard-loom weaves flowers and leaves.” That’s a quote by mathematician Ada Lovelace about the first programmable computing machine, the Analytical Engine.
Our language shows even more connections: we discuss related topics in “threads”. Whether textile threads or digital content and people, links are created by “linking”. We “zip” and “unzip” files, jackets, and bags. These parallels go beyond wordplay. To make the Analytical Engine programmable in the 19th century, its inventor, Charles Babbage, used a binary punch card system that he adapted from a different invention: the Jacquard loom.
So the digital world and textile craft already intersected at the beginning, when programming itself took its first steps in the textile industry.

How did your understanding of creative processes and the potential of these technologies evolve during the project?

Max: I would describe myself as a technology-positive person. So when I started this project in 2019, I had a very optimistic attitude about using AI in design. Prototyping is a crucial part of the design process, and AI can do a great job here, opening up a broad range of possibilities. This could shift the role of designers more towards that of a curator rather than a creator.

Today, however, I’m far less optimistic. Looking around, I get the sense that many people don’t use AI tools to enhance their creative process but rather hand their thinking and creativity over to AI entirely. The results are often not critically evaluated. This leads to a proliferation of “AI slop”. If this trend continues, AI may not only become the generator but also the curator. And I’m not sure that’s a good thing for the future of design outcomes.

Ida: My perspective has been more confirmed than changed. AI is a fascinating tool, but one you have to learn to use properly. There are aspects of artistic practice where AI can be used wonderfully. But there are others where it simply can’t. I love working with materials and sensing them. I feel the resistance of my pencils on the paper. I feel the cold clay and smell its earthy dampness… Sitting in front of a computer and typing commands doesn’t fulfil me. When its use arises naturally from the process or, as in MLrug, the central question, I can see it working well. But it won’t become my new “go-to” tool.