Generating Fabric Patterns from Text, in 2021

This is the oldest thing on this site, and the reason the rest of it exists.

In my second year of computer engineering at Bursa Technical University, a proposal of mine came back funded: TÜBİTAK 2209-B, the industry-oriented research programme for undergraduates, 2020 first period. I led it, Mert Bozkır came on as researcher, and the title was longer than most of the code — Doğal Dil İşleme ve Derin Öğrenme Yöntemleri Kullanılarak Metinden Kumaş Deseni Üretilmesi. Generating fabric patterns from text with natural language processing and deep learning.

You describe a textile pattern in a sentence. Colour, shape, size, how the motifs sit relative to each other. The system gives you back an original pattern that did not exist before.

Why this was a strange thing to attempt then

The proposal went in during 2020. If you want to explain this project today you say "text to image" and everyone nods. At the time that phrase did not mean anything to most people — OpenAI announced DALL·E in January 2021, while we were already working, and you could not use it anyway. Stable Diffusion was a year and a half out. Diffusion models were a research curiosity, not something you built on.

So text-to-image meant GANs, and getting a GAN to listen to a sentence meant building the bridge yourself.

The two halves

The NLP half is the part people skip when they explain this. A neural network cannot read "small navy paisley on cream, evenly spaced". It needs a vector. So the first job is representation: parse the description into the attributes that actually vary — colour, motif, scale, layout — and encode them into a fixed-length vector the generator can take as input. Get this wrong and nothing downstream can recover, because two sentences that mean different things end up as nearly the same numbers.

The GAN half is two networks arguing. A generator takes random noise and produces an image. A discriminator looks at an image and judges whether it is real or generated. They train against each other: the generator tries to fool the discriminator, the discriminator tries not to be fooled, and the pressure between them is what pushes the generator toward images that look like the training set.

The trick that makes it text-to-image is conditioning. You hand the text vector to both networks. The generator produces from noise plus text. The discriminator no longer answers one question but two: is this a real pattern, and does it match this description? A generator that makes beautiful patterns unrelated to the sentence now fails, which is exactly the pressure you want.

Why fabric was the right problem

Not because it was easy. Because it was narrow.

General text-to-image in 2021 was out of reach for two undergraduates with a university GPU lab. Textile patterns are a much smaller world: they repeat, they tile, they are built from a limited vocabulary of motifs, and the space of "correct" outputs is constrained in ways a photograph of anything is not. A model with a narrow domain and a clear notion of what a good output looks like can work at a scale where a general one cannot.

That is a lesson I have used constantly since: scope the domain until the problem fits the resources you actually have, rather than picking the impressive version and running out of runway.

The industry side mattered for the same reason. The project ran with Bursalı Tekstil, and Bursa is a textile city, so the question of whether a pattern was usable had someone in the room who could answer it. Academic advice came from Gizem Ortaç, industry advice from Hatice Kurt at Bursalı, and our department head Turgay Tugay Bilgin backed the thing before it had anything to show. We filed a patent application, ran it on the university's high-performance machines, and it ended up on Anadolu Ajansı — a strange experience when you are twenty and your model still sometimes outputs mud.

What it left me with

The system is old now. The approach is obsolete — anyone doing this today reaches for a diffusion model and gets better results in an afternoon.

What is not obsolete is what it forced me to learn. Conditioning a generative model on text, and the idea that the way you represent the request determines the ceiling of what comes back, is the same shape as the work I do now. Prompt structure, retrieval context, tool schemas — all of it is the same question in different clothing: how do you turn what a human wants into something a model can actually be steered by.

I did not know that at the time. I just wanted to see if the thing would draw.