MLRUG

I started this project a couple of years before many of the impressive AI tools (Stable Diffusion, VEO, GPT-5, etc.) we have today existed. The only realistic way for me to conduct this experiment was to train my own algorithm (based on other people’s algorithms, of course).

As with so many machine learning projects, the most important consideration was what data to feed into the algorithm. I needed a design object I was familiar with, for which I could get enough images, and that could be represented in two dimensions (since 3D generative design is much more difficult). Moroccan carpets were the natural choice. My father trades with them, which gives me access to a high-quality photo archive. I supplemented his collection with images from the internet, collected using a web crawler. In the end I had about 3.000 images. Not a huge amount in the context of a generative algorithm, but enough for a first experiment.

Even with a limited dataset and a relatively simple architecture, the results were instructive. The algorithm had already picked up on the diamond motif, one of the most characteristic and recurring elements in Moroccan carpet design. But the outputs had the uncanny valley quality typical of early image generators. Things that looked plausible at a distance but fell apart under closer inspection.

This taught me something useful about AI as a design tool. Even at this early stage, the algorithm was identifying the most dominant visual patterns in the data. But it couldn’t yet combine them in ways that were compositionally interesting or actually usable. It was clear that generative AI at this scale wasn’t going to replace a designer’s judgment. But it could serve as a source of raw material. Something to curate from, not to use directly. That distinction, designer as curator rather than creator, became the operating principle for the rest of the project.

For this second iteration I reevaluated which algorithm to use and landed on StyleGAN2-ada, a variant specifically designed for training with fewer images. Using my father’s high-quality photo archive again, this time supplemented by much more carefully selected internet images, we achieved dramatically better results. The leap in quality demonstrated two things: how rapidly machine learning had evolved in just a few years, and how much of a difference thoughtful data curation makes. Good data isn’t just more data. It’s data selected with understanding and intent.

Producing the carpets in Morocco was the obvious choice. We had trained on Moroccan carpets, so bringing them to life through traditional Moroccan production wasn’t just appropriate, it was the only approach that made sense.

The alternative would have been Indian or Nepalese manufactories, where machine-assisted production can reproduce a design nearly knot for knot. The result is a faithful reproduction, but it lacks the character of the traditional Moroccan approach. In the workshops of the western Middle Atlas, where my father has collaborated with a producer for several years, things work differently. The process is low-tech. We sent our generated designs as image files, which the workshop printed on a local printer onto A4 pages. The weavers work from these printouts by eye, interpreting the design rather than mechanically reproducing it.

This means every rug is unique, shaped by the weaver’s interpretation. What gets “lost” in translation is exactly what makes the result interesting. The weaver’s hand introduces the same kind of improvisational quality that has always characterized Moroccan carpets. And there’s something conceptually satisfying about it too. The pattern has now been transformed yet again. It migrated from historical carpets into a training dataset, was recombined by an algorithm, curated by us, and then reinterpreted by a human weaver. Each step is another layer in an ongoing process of pattern transformation, not unlike the centuries of cultural migration that created the source material in the first place.

These algorithmically generated, curated and handcrafted rugs were presented as part of Design Month Graz 2025.

The reception was more positive than I had expected. There is no shortage of AI-generated work out there right now, and a lot of it provokes skepticism, especially when it encroaches on traditional craft and art. But the response to this project was genuinely warm. I’d like to think that’s because the work was done with enough care and intention to not come across as AI slop. A purpose-built dataset, traditional production, and a clear conceptual framework seem to have helped.

The project was featured in two articles in Cover, a leading publication in handmade carpets and textiles for interiors. Lucy Upward’s feature article^[1] explored the project’s conceptual framework, examining how AI’s pattern-learning process mirrors the historical migration of motifs along ancient trade routes. Denna Jones’s analysis^[2] used MLRug as a case study for ethical AI implementation in design, by keeping production in Morocco and creating a purpose-built dataset to avoid copyright concerns.