The Last Saree: Connoisseurship in the Age of AI

Is the "era of saree draping silently coming to an end"? Hidden inside the clickbaity title of both this article and the original question on Shilpa lies a deeper anxiety. What if, as we careen forward in embracing new technologies, we forget the heritage that makes life colorful, diverse, and enjoyable? Rejecting the Luddite's anti-technology violence, can we envision a world where technology might be put into the service of preserving artisan material culture, rather than erasing it? In this article, let's explore an intersection of technology and artisanship by studying the special case of the artisan textile as a literal web that connects old and new, artisan and engineer, poet and scientist.

Textiles as Catalyst

Saree-style draped garments have been depicted in carvings over 1000 years old. It's important at the outset of a thought piece such as this to understand that ancient traditions can never be 'improved' by modern technology. It is debatable whether power looms, for example, can produce textiles comparable to those handcrafted on handlooms.

A Terracotta figurine draped in a saree-like garment from Bengal (200–100 BCE)

Textiles, as aspects of culture that are relevant to almost every human on earth, are uniquely interesting to study as artifacts that form webs of important interpersonal and intercultural connections. Artisan textiles have been under threat since the industrial revolution, when some of the first advances in industrial technology were used to mechanize processing raw weavable materials and creating cloth. Textiles, thus, are at the nexus of the tension between technological advances and the desire to preserve artisanship. They can be fruitfully used as a case study to consider the effect that advances in AI might have on material culture.

Moving from Cultural Appropriation to Cultural Appreciation

The study of textiles often leads to the study of fashion, and how humans use clothing to signal affiliation. Sometimes, the inappropriate or thoughtless use of a garment used by a group signals ill-intent, or simply ignorance. Wearing native American war headdresses at football games, for example, is taking a garment with important cultural significance and wearing it outside of its appropriate context. If we presume that the application of technological advances can be put into the service of moving an audience from incomprehension to understanding (the premise of EdTech), then let's observe how we might use these tools in service of helping an audience gain awareness of the glorious artisan textiles of India.

Textile as Story

One way to battle ignorance is by observing history. Where did a given element of material culture come from, and how did it evolve to its current form? What cultural contexts informed that evolution? A good example of a garment that evolved over hundreds of years is the good, old-fashioned, blue jean...that quintessential American garment (or is it?)

The story of jeans starts around the 17th Century when blue pants originated in India and were worn by sailors - these are called Dungarees possibly from the Indian town of Dongri. They may be the precursor to "Serge de Nîmes", a fabric inspired by Indian cotton and the origin of the word 'Denim'. The word 'jeans' originates from a type of material made in Genoa - 'Bleu de Gênes'. Note the association of these artifacts to port cities where dockworkers and sailors need good, strong working pants, as well as the rapid movement of this garment around the world by being associated with voyaging sailors.

The indigo-dyed twill that blue jeans became is strongly associated, in the Americas, with forced labor on Southern plantations and the movement of cotton between North and South. The fabric can be said to have profoundly influenced the history of the United States. The blue jeans that we know today were canonized by the Levi Strauss company which added, and patented, the use of rivets on the pants' seams in the 1870s so that miners in California could depend on their work clothes not to rip. By the 1970s jeans had become a symbol of the counter-culture, by the 1980s high fashion, and they continue to evolve today. Following the history of the humble blue jean allows us to equate the evolution of a textile to the cultural currents that carried it around the world.

The indestructible Levi-Strauss blue jeans!

Technology for Storytelling

Where does technology fit in to looking at an artifact's history? Museums around the world are grappling with this very question today. The first step in this direction involves bringing a collection online, and to do that it needs to be digitized.

DAM (Digital Asset Management) solutions are used by museums and libraries (among other business entities) to digitize an analog collection. All the photographs, 3D renderings, and metadata that can be stored in an online system can be well served by a robust DAM system. The APIs built by the Cooper Hewitt and Metropolitan Museums, for example, are fed by a DAM system coupled with an open-access software license so that developers can easily interact with a collection.

The simplest DAM system would be a relational database able to store multiple photos and renderings of an object alongside its metadata including provenance, medium, date, acquisition date, and more. A well designed database of this type would allow users to query related artifacts across departments, connecting, for example, a porcelain Ming dynasty vase with a work on paper from the same era, or given by the same donor. But more and more complex technology is being used to assist both in the cataloguing process. OCR (Optical Character Recognition) and Natural Language recognition for example can be used to decipher typed and handwritten labels in a specimen set.

I have shown elsewhere how machine learning can help to identify items in a collection, given enough quality training data. Using standard machine learning training techniques, we are able to create a ML model to aid in the categorization of a set of artisan textiles. The steps to follow are straightforward:

• Gather acceptable-use images from the web
• Categorize them into labeled folders
• Drop them into a storage system (such as S3):

• Use a service such as Amazon Rekognition or Azure AI Vision to perform transfer learning on your dataset and create a custom ML model:

With a small number of images (I found 130 images of saree fabric styles, categorized by location and saree weaving type), you can train a testable ML model.

Once your model performs to your satisfaction, you can call it in a web app. The full codebase is here and you can watch a demo here:

A demo of the Saree Detector app. Note how the images are determined with surprising accuracy.

A basic use case of a web app calling a custom ML model, then, provides an example of a way that someone managing a collection could use computer vision as a way to triage items, at least a first pass.

ML to AI

What about the use of AI, and specifically generative AI, to support the appreciation of cultural heritage? I was lucky enough to travel to India's Gujarat region and visit the Patola weavers of Patan. These are artisans using handloom techniques and a way of tie-dying both the warp and the weft of the fabric threads before it is woven, producing single and double ikat fabric, in what seems almost like pixel art, or a technique similar to creating pysanka eggs.

A saree weaver in Patan, Gujarat, India shows the tie-dye process of a Patola style saree weave. Photo by the author.

It's important to realize that no AI process can ever improve this art form. It has been handed down over generations and we should see it as something to treasure and preserve. So it is fruitless to wonder whether Generative AI, which creates something novel based on images it has already ingested, can create something better than a Patola saree. Instead, we should focus our attention on whether there are sufficient quality images available on which to train these models, which takes us back to DAMS.

AI Era Risks and Promises

Artisanship might be seen as being at risk during this brave new era of AI. The content created by these systems is by definition disconnected from its source and attribution can be muddled. Results can be superficial or even hallucinatory. Source can even be erased. At the same time, the wide availability of systems such as ChatGPT and PartyRock.aws have dramatically democratized some very high-tech systems, opening the door to more diverse voices to participate, which will improve the results generated.

I tested a Generative AI system (PartyRock.aws) to see how far I could push it to generate a saree design in a given style by building SilkWeave, an AI app (try it yourself!). The results were mixed:

PartyRock-generated sarees

I pushed it to deliver several iterations of a Cloud design, however:

I was pleased enough with the result to commission it to be embroidered for me to wear in India: the world's first AI Saree!

"SarAI" - the world's first AI generated saree, purple silk with resham threadwork clouds, stitched in Gujarat.

Diving Deeper: Towards Deep Culture

Artisan textiles, however, have deeper meaning than is often recognized. The motifs for these expensive garments are often carefully chosen and have meaning. Banarasi weavers, for example, employ a visual vocabulary that bears careful study. Could we ask a generative AI system to create for us a saree design based on the meaning that we want the motifs to convey? A sort of secret code?

Banarasi saree motifs have a whole vocabulary of paisley, floral trellises, mangoes, creeping vines, hunting scenes, and more. Asking an LLM to generate a saree with the following prompt "Banarasi Saree design symbolizing good luck, fertility, and prosperity via the symbols in the background and pallu" generated a mess:

Clearly, the model has no exposure to this type of very specific and deep culture. However, there is a way to make an LLM a little more precise: by using a Knowledge Base. You can find them connected to Amazon Bedrock here and they act as vetted sources that can improve your prompt.

This knowledge base is built by connecting an LLM to vetted online sources.

With a much better prompt, a better image can be created:

This design symbolizes prosperity, good luck and fertility by including mangoes, paisley, floral jaal, and bulbul birds as suggested by the enhanced prompt

In conclusion, the use of technology to safeguard and better understand material culture leads inevitably to today's AI moment. Approached carefully, we can use AI not to generate new artifacts stripped of meaning and provenance, but to better understand and appreciate the heritages that we inherit, preserving them for another generation to cherish.

This article was originally presented as a TEDx talk at CHARUSAT university in Gujarat, India

Slides from the talk can be found on Slideshare