The intricate historic interweaving of textile manufacturing and digital technologies opens unlimited transformational opportunities at the dawn of the new era of artificial intelligence. While fashion and textile industries already leverage AI-powered tools for real-to-virtual transformation of products and processes, we make the case that AI can and should play a key role in enhancing virtual-to-real product transformation via generative design of textiles for manufacture.
Although generative AI will inevitably impact workforce development, we are optimistic that it will also provide the designers, artisans, and hobbyists with new tools to preserve and elevate their craft. The combination of the language of words and textile patterns can become a powerful communication tool, most famously practiced by late Justice Ruth Bader Ginsburg through the sharp use of language punctuated by ornate lace collars she wore on the bench to communicate her opinions or dissents.
Just like languages, traditional textile crafts offer a tremendous opportunity for AI-enabled generation of new knowledge, thus closing the loop between knowledge generation and storage and textile engineering. We propose and discuss a generative AI-enabled pipeline to design textiles for manufacture, which integrates historical pattern collection studies, mathematical modeling, mechanical characterization, computer vision deep learning, and lacemaking knowledge.
While our pipeline is applicable to any textile type, we focus on bobbin lace as an intricate, challenging example of an endangered handicraft important to textile heritage. Lacemaking draws from elements of weaving, embroidering/sewing, and knitting, all current methods of mass textile manufacture, but adds the challenge of holes and negative space in the fabric, forming an intricate pattern. Relative to more conventional woven or knitted textiles, the open net structure of lace textiles provides additional degrees of freedom in tensile properties engineering, which can be leveraged for modern applications in wearable, medical, industrial, and geo textiles.
AI-generated lace patterns can be optimized for aesthetic appeal, cultural relevance, elasticity, tensile strength, Poisson ratio, and other mechanical characteristics as well as for the integration of conductive threads and electronic components. Historically, lace has been an indicator of wealth, class, and decorative flair but has fallen out of popularity in modern fashion. Lacemaking was once a valuable source of income for poor women, skilled artisans who advanced the craft but often remained uncredited for their artistry and engineering skills.
The cheap mass production of chemical lace and bobbinet machines led to a decline in handmade lacemaking and devalued the intricate craftsmanship and cultural significance of traditional techniques. Nowadays, designers, artists, and writers fear similar displacement by generative AI technologies. Here, we make the case that generative AI can bridge the gap between historical craftsmanship and contemporary technology to create a sustainable model for the preservation and evolution of textile heritage crafts. If leveraged properly, AI can balance the valuable contributions of historical and current lacemaking practitioners to enable the revival of the lacemaking craft not merely as a historical relic but as a living, evolving art form that adapts to the needs and aesthetics of the present.
Traditionally, lace has been a culturally significant commodity. It was used by European nobility in the seventeenth and eighteenth centuries to display social status, wealth, and fashion trends. Lacemaking was typically practiced by women and children, providing means to contribute to their household finances or a dowry, make money of their own, and have a skilled trade that they could take pride in. The industrial revolution brought machinery to the world of lacemaking and changed the craft. While before, lacemaking was primarily done in houses by groups of women, it now moved to factories, leading to an increase in productivity at the expense of shuttering down small-scale handmade production.
Nevertheless, nostalgia and longing for the handmade craftsmanship of previous times helped to preserve the practice of traditional handmade lacemaking. Despite market changes and demands, lacemaking styles evolved overtime to become pillars of practicing communities’ culture and heritage and played a major role in cultural exchange between countries and continents. Having originated and flourished in Europe, lacemaking craft was inspired by intricate patterns found in Middle Eastern and Asian woven textiles. Designing and preserving lacemaking patterns has always held deep value and importance, from the noble women of 1600s Venice, Italy, boasting books containing needlepoint lace patterns from all over Europe to twenty-first century women in Central Slovakia taking great pride in and marketing their pattern collections.
Further evolution and fusion of lacemaking styles were shaped by immigrants from different parts of Europe moving to new countries, bringing their lacemaking practices with them, and ultimately forming a new style as they mixed their craft with those of the new communities. Unfortunately, some unique lacemaking techniques and artforms are endangered or vanishing, such as Spanier Arbeit, a unique metal wire-based two- and three-dimensional bobbin lace exclusive to Ashkenazi Jewish production, which was decimated by the Holocaust.
Even within actively-practicing communities, pattern heritage preservation is not consistently maintained. As an example, within the Central Slovakian lacemaking community, the older and younger generations have diverging views on lacemaking instruction and pattern preservation. The older generations view lacemaking as a craft to be learned directly from a master through examples and practice with direct supervision until the pupil develops an intuition for the craft. Many older lace makers do not keep their patterns after using them, sometimes even burning patterns they do not like—a practice that is viewed as irresponsible by younger craftsmen who wish to preserve the cultural knowledge embedded in patterns made by master lace makers.
For an AI model to produce new lacemaking patterns, it must train on a large database of existing historical patterns. Given the age of the lacemaking craft, copyright and intellectual property issues should not be a barrier to access and use training data for generative AI models. Training on historical lacemaking books in the public domain will enable the generation of new digital design-for-manufacture tools while benefiting the craft by preserving lace patterns from being lost to time with aging older generations and declining interest.
As generative AI will inevitably impact industries and workforce development, it is also important to consider what artisans and cultural heritage may be displaced by new technology or, to the contrary, provided with new tools to preserve and elevate the lacemaking craft. Like many industries, the fashion industry in general seeks to reap the benefits of AI technology. Big data surrounds fashion, and companies hope to use this information to train generative models to create relevant and practical designs. Generative models are compelling for their ability to take advantage of real-time data and trends, personalize output to user preferences, and accelerate productivity.
Historically, interactive genetic algorithms (IGA) have been adopted to inform the computer-aided design of fashion and evolve designs based on previous ones. While conventional genetic algorithms emulate natural selection and solve optimization problems by evaluating the fitness of each candidate solution against a predefined objective function, IGA allows the fitness function to be chosen by the user. However, these methods are still limited because they rely on discrete features such as style, silhouette, color, and pattern instead of using the overall design composition or microscopic textile structure as an input.
In turn, neural networks have been used in the creative design process to pair fashion with human emotion, to recommend fashion looks based on previous user preferences, or to perform textile classification. Generative adversarial networks (GANs) created by Goodfellow et al. in 2014 set the tone for deep generative AI and opened new horizons in fashion design applications. GAN is an unsupervised deep learning architecture, which generates new data via two neural network components, the generator and discriminator. The generator creates the new data samples (e.g., images), which are then mixed with real data samples and run through the discriminator tasked with correctly classifying these sample types.
Since their creation, GANs have already had an impact in generating new fashion designs and have been recently joined by variational autoencoders (VAEs) and diffusion models. Each of these deep learning methods encodes stylistic information into a lower-dimensional latent space and represents different styles as a probability distribution, which can be sampled. Deep learning has also aided in textile visualization using image-to-image transfer, specifically neural style transfer, a technique commonly used in computer vision and machine learning in general. The process translates the semantic content of data of one domain to another.
Recently, StyleGAN was developed by a group of NVIDIA researchers with the goal of improving GAN to generate highly realistic images. StyleGAN introduces an intermediate latent space that controls “styles” of the generated image during the image generation process, such as textile texture, pattern design, and gradient. This allows the model to have precise control over various aspects of the image, which the model adjusts as it learns to produce the most realistic image possible. Furthermore, the model also incorporates other techniques that improve the realism and resolution of the model’s outputs, such as progressive growing, mixed multiresolution technique, and noise inputs in the generator model.
As a result, it generates high-quality images that exhibit a diverse range of appearances and structures. Due to these attributes, StyleGAN excels in creating detailed textile images, despite not specializing in textile generation. In a trial survey of StyleGAN involving 200 users, knitted textile images were translated into swatches, and it was found that these swatches “were rated overall more creative, fashionable, and buyable than ones based on the real knitted textile images”. The later version, StyleGAN2, improves upon the realism and resolution of StyleGAN’s outputs by removing artifacts and addressing some limitations.
Such methods would be especially useful as built-in digital tools for fashion designers depending on visual specifications such as color or stitch type. Entire libraries of new patterns can be quickly generated with these deep learning models. So far, generative AI models have worked on image generation for textile patterns and have not yet fully evolved to the level of creating reproducible textiles. Despite advancements in knitted pattern generation and even symmetrical lace pattern generation in SStyleGAN, the current phase of AI-generated content typically takes the form of pieced-together images that are not yet manufacturable.
In this context, we define manufacturability as the ability of an AI model to generate a set of instructions sufficient to produce a textile from a generated pattern. For AI-generated textiles, the challenge lies not only in creating a generative AI model capable of producing patterns on demand but also in ensuring patterns are complete, physically possible to be made, and encoded to be made by hand or machine. Accordingly, we identify three critical stages in the AI-enabled textile design-for-manufacture process: attribute-specific pattern generation, process-specific instructions encoding, and physical fabrication.
The attributes may comprise aesthetic as well as textural, mechanical, or thermal features, while process-specific instructions can take many different forms, depending on the choice of the manufacturing process (i.e., weaving, knitting, bobbin lacemaking, three-dimensional printing, or embroidery). Modern knitting is a fully computerized textile construction technique, which creates patterns from interlacing yarn loops comprising various stitch types (e.g., knit, purl, tuck, flow, etc.) used as pixels. Punch cards, originally created for the Jacquard loom, have provided a way of encoding lace in a binary way while retaining the ability to make intricate designs.
They have been used for the production of different lace styles such as Fair Isle, punch lace, knit weave, or tuck stitch. Fair Isle knitting, for instance, is a method used to create patterns with multiple colors via a stranding technique (process of moving strands along the back of a work). In this example, the binary of the punch card pattern represents the assorted colors of the lace. Fair Isle knit designs are typically seen as pixelated designs that are repeated and mirrored to form patterns. In turn, the tuck lace technique allows for more geometric diversity in the lace topology since it allows for holes to open and the density of the overall design to vary.
AI pattern generation from punch card images by GAN and neural style transfer techniques has been used to produce Fair Isle knitted laces. Machine knitting has been recently revolutionized by several practices, including (i) whole garment knitting, which enables seamless three-dimensional garment manufacturing, (ii) assembly of knit primitives (such as tubes, sheets) into low-level machine instructions, and (iii) pipelines to generate new patterns from pre-existing ones. These innovations simplify assembly of complex stitches, but do not generate new patterns.
Differently from knitting, bobbin lace is made by braiding and twisting filaments or yarns, which are wound on multiple bobbins. Simple movements of the bobbins (e.g., twists and crosses) create stitches according to a predefined pattern. Manual bobbin lace technique uses patterns drawn on paper or parchment and pinned to a lace pillow, where the placement of the pins determines the pathway for the lace stitches. Handcrafted bobbin lace instruction typically includes both an encoded sequence of stitches and a visual model of pricking patterns, subject to interpretation on what stitches to do at each point of intersection, and checking completion is a critical evaluation when determining pattern feasibility.
The bobbin lace technique has been mechanized and digitized for mass production by the invention of the Leavers loom in 1813. John Lever adapted a Jacquard loom head for use with the bobbin net machine engineered by John Heathcote in 1808, allowing for complex lace to be machine made by imitating the basic movements of the handmade technique. Leavers looms produce lace by intertwining two sets of threads: (i) the warp and beam threads that are actuated by the Jacquard mechanism move right or left and (ii) the bobbin threads, which always move along the same path as the bobbins, swing back and forth in a pendulum-like motion. The patterns created by the loom can be digitized by using a binary code with punch cards or computer codes.
An alternative lace encoding approach has been recently proposed to represent these patterns as graphs, which allows effectively integrating bobbin lace patterns as quantifiable data representations. Bobbin lace patterns can be represented as simple graphs known as grounds, in which nodes represent an encoding of actions (twist, cross, etc.) to be done at said node, and edges represent topological threads between each lace. Each combination of actions and threads can be represented by a special syntax. By representing these bobbin lace patterns as graphs, we can use state-of-the-art generative AI models, such as graph neural networks, to further explore and innovate bobbin lace textile designs.
Graph neural networks particularly excel at processing graph data, allowing them to capture complex features and relationships between edges and nodes. As a result, they can generate not only realistic textile graphs that reflect these relationships but also reproducible instructions to create them. This is not something image generation models can achieve, as they do not have actual stitch data. The adaptation of three-dimensional (3D) printing techniques for lace manufacture faces a completely different challenge. While a sophisticated and flexible system of coding 3D patterns is well developed and standardized, reaching the same level of material flexibility, aesthetics, and production rate as those achieved with knitted or bobbin lace remains a challenge.
On the other hand, different additive manufacturing techniques—including selective laser sintering, fused deposition modeling, and two-photon lithography—make use of the well-developed software tools such as Rhinoceros and Autodesk 123D 3D computer-aided design (CAD) modeling tools. CAD files are further converted into either the stereolithography or the additive manufacturing file format by tiling or tessellating the models’ geometrical surfaces with triangles. As a result, a 3D model is converted into a geometric (G)-code—a machine instruction file with the instructions of the print path—and then into physical patterns via 3D printing techniques.
Using the methodology of G-code to generate lace structures also opens the opportunity of constructing 3D interlocking lace patterns using digital embroidery. In this process, repeated patterns produced by a computer-controlled embroidery machine in layers can build up new interlocking mechanisms. While each layer has its own mechanical properties, the construction creates joint spots that interlock between the layers for the creation of 3D lace when the dissolvable backing fabric is removed. The embroidered lace process is similar to chemical lace production but allows for increased complexity by using a 3D layering technique. In contrast to bobbin lace, embroidery—which traces its origins to the needlepoint lace techniques—uses only two continuous threads per layer, which allows it to be produced on more commonly available machines.
Our process exploration invites the potential to embed an AI system into existing G-code methodologies to generate layer images for these lace structures with unique properties and intricate overlaid patterns. While aesthetic properties of lace structures may be driving their consumer appeal, it is the mechanical properties that play a critical role in determining the suitability of different patterns for various applications. In this context, key attributes include tensile strength, elasticity, dimensional stability, fineness, and texture. The consideration for each of these attributes allows for the creation of complex lace patterns, including structures that can bear distributed loads or move in interesting ways like auxetics in metamaterials.
Similarly, creating lace that can deform either uniformly or nonuniformly in a predesigned fashion gives value to wearables or other scaffold structures. Generative AI is expected to expand the possibilities of incorporating these attributes and nontraditional materials in lacemaking practice to both elevate the craft to be of research significance and bring functional and aesthetic value to modern textiles. When developing advanced generative AI models under the design-for-manufacture paradigm, tensile properties of
