What AI can (and can’t) do for rug design

Will artificial intelligence be a field of dreams or a minefield for rug designers? In our second segment of the AI series, Denna Jones leads us cautiously forward

Quality and control. Remember these words if you plan to join the AI (artificial intelligence) FOMO stampede to design your next rug collection.

Generative AI (GenAI) is currently the most advanced iteration of AI for image content creation. COVER editor Lucy Upward’s AI talk at Heimtextil 2025 generated audience questions on how companies can use AI to design rugs, but there is no single or simple answer.

If you want GenAI to help you brainstorm or create detailed rug designs, then budget, employee time, and tech skills considerations will lead you to choose one of several GenAI programs, or to the decision to create your own bespoke AI program.

‘Modern Technology for an Ancient Craft’ is the strapline of a well-known custom carpet design software. This type of software relies on rug pattern input that is then refined and readied for the weavers. GenAI, however, freestyles a pattern or image based on a written prompt typed into the program by the user. The image is created using datasets trained on billions of images scraped from the internet.

High-quality output from GenAI depends on an extensive library (i.e. dataset) of genuinely high-quality images. The program is of little use if a tasker captions an image ‘grandma’s rug’ because they don’t know the rug is an antique Persian Garden design Kerman rug. Also, the program needs a large number of the type of images that will be useful to the user. It’s unlikely even the largest GenAI programs are stuffed full of rug images, which is one reason (but not the sole reason) why one company— Neuroweave—and one non-commercial research art project—MLRug—created bespoke AI datasets dedicated to specific categories of rug images.

There’s a tech term called ‘garbage in, garbage out’, which basically means the computer is only as good as its user. Companies are hiring prompt engineers, and here’s why: an advanced prompt uses symbols, characters or punctuation marks to add clarity, control and context so that GenAI truly ‘understands’ you, because (for the moment at least) GenAI is not a mind reader.

AI and GenAI legal issues are another reason why Neuroweave and MLRug created their own datasets. We are in a cultural lag where existing laws are not fit for purpose. There are multiple ongoing court cases with creatives suing AI companies that scraped their work and used it without permission to train AI. If a prompt uses work by living artists and designers, or works where multi-decade copyright rules apply, then, as laws evolve, the downstream user could be held liable for rights infringement.

Neuroweave co-founder Mehmet Bozatli describes how ‘our dataset was built entirely from my own archive of traditional Persian and oriental rug images photographed and catalogued over two decades’. Neuroweave’s bespoke dataset avoids infringement concerns ‘while engaging with a design tradition that has long evolved through adaptation. That said, we believe ethical considerations around IP and fair compensation for creators must be central to the conversation around AI.’

Similarly, MLRug (see Lucy Upward’s article in this issue), created by Ida Hausner and her brother Max Blazek, produced thirteen rugs woven from designs generated by a dataset. The dataset was trained on approximately 1,000 hi-res photos of Moroccan rugs from their father Gebhart Blazek’s archive, plus more images from the internet. They used StyleGAN2-ada: a GenAI program designed to train effectively with fewer images.

The rugs were woven in Morocco to keep production in country of origin, but also, Gebhart says, because rugs in Morocco ‘are much more based on improvisation than in other countries’. Improvisation mirrors the character of handmade, but it also reflects the project’s focus at how AI learns and copies, compared with historic human-tohuman analogue evolution of rug patterns.

An AI takeaway for 2025 is awareness that writing a prompt that includes ‘in the style of’ and then adding the name of a living rug designer tempts fate. copyright and laws often impact global law. For the moment, US copyright protects original artist works, but not their ‘style’.

However, the US Copyright Act of 1976 has a ‘useful article’ limitation that states that, although copyright generally does not protect functional objects like rugs, it does protect artistic expression if the design within the functional object incorporates pictorial, graphic or sculptural features that can exist independently of the object. This separability distinction was supported by a US Supreme Court ruling. The simple protection device for designers may lie in creating a painting or drawing of the artwork that you use for a rug design.

Max Blazek was optimistic about the future of AI in 2019 when he began his project. In 2025 he’s not so sure. ‘Looking around, I get the sense that many people don’t use AI tools to enhance their creative process, but rather to hand their thinking and creativity over to AI entirely.’ Like most tools, GenAI is neutral. Whether it evolves into a tool that assists rather than replaces human creativity is up to us.

These are non-comprehensive steps for designing with GenAI or creating a bespoke dataset. The Tufting Shop, Netherlands, produced a GenAI guide for creating hand-tufted rug designs. We have lightly edited their steps:

1. Use DALL-E, Midjourney or Stable Diffusion.

2. Suggested prompts: Cartoonish figures, abstract patterns, mid-century modern.

3. Convert design to graphic format using Adobe Firefly, Runway ML or manually in Adobe Photoshop or Illustrator. Clean design and adjust colours.

4. Abstract the design for tufting. In tufting, stitch resolution is typically measured as stitches per inch. ‘Abstract the design’ reflects the fact that tufting is best suitable for clear, simple shapes and solid colour fields.

5. Transfer the design to canvas for tufting.

A suggested baseline for a dataset is 1,000 images for each classification (e.g. Kashmir). Balance the quantity of images with the quality of images to optimise training and performance.

1. Gather, clean, and crop images (e.g. a living room with a rug = crop to rug only).

2. Label images according to their classification.

3. Move images into one folder.

4. Resize images to uniform dimensions.

5. Convert images into the same file format: png and jpg are the most common for AI datasets.

6. Convert images into a plain text Comma-Separated Values (CSV) file.

7. Tweak/edit the CSV file.

8. Load the CSV file.