The Straw Hat That Sells in 70 Countries Is Being Remade by AI

A while back, I came across a study by a research team.

Guess what their research subject was.

Straw hats.

The kind you wear at the beach in summer, the kind that runs you a few dozen yuan apiece. The team went to Cixi, Zhejiang, and embedded themselves in the local straw-hat industrial cluster, trying to answer one question: how do you run a cross-border business with data and AI, on the most traditional, least high-tech small commodity imaginable?

The more I read, the more interesting it got. Because the story of this straw hat is, in miniature, the story of every small Chinese manufacturer.

First, an Invoice That Hurts

Some background first.

Woven straw hats are an old craft in China — braided from water grass, wheat straw, palm rope. Eco-friendly, breathable, an intangible cultural heritage going back thousands of years. Zhejiang is the heart of the industry: Cixi alone, a single county-level city, has more than 50 straw-hat factories with an annual output of 250 million yuan. The gold-thread straw hats made in Cixi's Changhe Town sell to nearly 70 countries and regions worldwide.

Sounds like a decent business, right?

But at home, there's overcapacity, demand is saturated, and the price war has hit a ceiling. More and more factories are looking overseas, turning to cross-border e-commerce.

And the moment they go abroad, they fall into two big traps.

The first trap: labor costs. As soon as a foreign-trade order comes in, every step — client liaison, design, sampling, contract signing — piles on people. More orders, more people, thinner margins. The researchers did the math: the hidden costs of all this back-and-forth communication and sample revision can account for 30 to 60 percent of the total cost of processing an order.

What does that mean?

Suppose an order costs 100,000 yuan to process. Up to 60,000 of that can burn away in "you say, I say" communication.

The second trap: not understanding overseas consumers. The name of the game in e-commerce is spotting selling points and engineering viral hits, but small factories have neither the technology nor the data. They produce on gut feeling, and the moment the wind shifts, they're stuck holding inventory.

So what do you do?

Let the Data Answer First: What Does the Market Actually Want?

The research team calls this the "surface problem": what does the market want, what do consumers want. That, data can dig out.

They started on Alibaba 1688, pulling the full-year 2023 wholesale data for "straw hat" and "woven hat": 3607 product records, 187 attribute fields, 49 left after cleaning. Then they went looking for patterns.

I'll pick the three most counterintuitive ones.

First: customization.

In the data, 72% of merchants offer customization — adding a logo, making to order from a customer's picture, gluing on rhinestones, doing camouflage, all of it counts. Guess who earns more, the customizers or the non-customizers?

Looking at average monthly sales revenue, customizing merchants pull 1198.68 versus 676.3 for the rest — nearly double. But wait: there are more than twice as many customizing merchants, the sample sizes are wildly different, and averages get skewed.

So they switched to the median: 16316.64 for customizers versus 13449.60 for the rest. Customization still wins. Even more interesting is the interquartile range (IQR, the spread of the middle half of the data): 61266.6 for customizers versus 68572.8 for the rest.

What does a smaller IQR mean? It means the income distribution is more concentrated, more stable, less easily hijacked by extreme values.

Customization isn't just a revenue add-on. It's insurance against the risk of dead stock.

Second: use scenarios.

87.4% of straw hats are tagged with a single use scenario — commuting is commuting, parties are parties. But the data shows that hats tagged as suitable for more than 5 scenarios have noticeably higher average sales. Same with seasons: the more seasons a hat fits, the better it sells; hats tagged "winter" actually scare buyers off — in consumers' minds, a straw hat means coolness and style, nothing to do with keeping warm.

In the e-commerce search box, versatility is traffic.

Third: price bands.

A hat wholesales for around 10 yuan domestically; on Amazon, it can sell for 10 to 100 dollars. The team scraped buyer reviews of bestselling straw hats on Amazon and mined them with an LDA topic model. Three price bands, three completely different crowds.

Under 25 dollars: people talk about size, whether it fits their head shape, whether it works as a gift, whether it travels well. Utility first.

25 to 50 dollars: they talk about whether the photos match the real thing, how customer service is, whether a solid color matches their wardrobe. This crowd is buying an experience.

Above 50 dollars: they talk about materials, craftsmanship, design, sustainability. This crowd isn't buying a hat — it's buying taste and a lifestyle.

Price bands are a sorting machine. The same category, at different price points, maps to three entirely different demands.

Now, if you combine these patterns, can you compute the "optimal straw hat" directly? Yes. Put a decision tree model on it; after training, test-set accuracy comes to 0.755. Rank each attribute's contribution, and the number one genuinely surprised me:

Not the hat shape, not the color. It's the store's service rating, with a contribution of 0.5857, far ahead of everything else (color 0.1361, style 0.1241).

Style decides whether customers want to buy. Service decides whether they trust you enough to hit the button.

The optimal recipe came out too: dome crown with no brim, original design or artistic vintage style, 5 or more color options, positioned for spring and summer, with the store's service rating pushed to around 0.9.

That's the data part — the surface problem, solved. But.

What's Really Expensive Is the "Foundational Problem"

Knowing what the market wants is one thing; you still have to make it, fast. The most valuable capability in foreign trade is response speed — the research team calls this the "foundational problem."

How does a traditional foreign-trade order run? Four steps: inquiry, quotation and sampling, contract and production, delivery and settlement.

Which step is the most expensive? Inquiry and sampling.

The client tosses over a design sketch, and the factory has to have designers and graphic artists produce sample previews, back and forth, confirm and reconfirm. Once production starts, one "change the color" from the client means the designer redoes the work and produces new images. Half a planet apart geographically, everything runs on online communication, and time just bleeds away. The bulk of that 30-to-60-percent hidden cost from earlier lives right here.

The research team's solution, stripped to one sentence: let AI play middleman.

Three changes.

First, clients no longer have to supply design requirements precise to the millimeter. The AI understands the requirement on its own and generates sample images directly; the supplier confirms against the images.

Second, mid-stream changes go from "designer reworks for days" to "images out in minutes." Change a color? Adjust a style? Nearly zero waiting.

Third, the factory builds its own order database. After all these years of selling, the accumulated historical requirements and product images are a gold mine. A new order comes in: first match against similar sample images in the order history, and only generate if there's no match.

The logic is smooth. But the moment it lands on one small factory, two very real questions pop up: which model do you use? And how do you make the model understand straw hats?

One Straw-Hat Factory's AI Makeover

The research team brought in the largest straw-hat manufacturer and exporter in Cixi — a factory called Jusheng — and ran the whole thing for real. This is the part I find most worth watching.

First, choosing the model. At the time, among the image-generating AIs on the market, the foreign options included Stability AI's Stable Diffusion (open source), OpenAI's DALL-E, and Midjourney; the domestic ones included Baidu's ERNIE and Alibaba's Luban. Each has its strengths: DALL-E is explosively creative and good with complex text descriptions; Midjourney leans abstract and distinctive, suited to film and game concept work; ERNIE and Luban understand Chinese and are cost-effective for e-commerce materials, but most of them aren't open source — you can only apply for API access.

A small factory's way of choosing is beautifully plain, just three rules: free (open source), aimed at overseas markets (pick an international model), and the data must never leave the building (local deployment).

There's only one answer: Stable Diffusion. Open source, free, downloadable to run offline on your own machines — not a single client data point leaks out.

But a general-purpose model used as-is has two pitfalls. One, its training data comes from murky sources, which creates legal exposure on copyright. Two, it can draw anything except the texture and style of your products. And clients at the inquiry stage are staring at exactly the details.

What to do? Fine-tuning.

Using the company's own data, give the model a low-resource fine-tuning (technically, the LoRA method): compress the parameters being trained to one ten-thousandth of the original, cut GPU memory usage by two-thirds, with no hit to inference speed. The training data was about 3,000 pairs of "order requirement text → final sample image," all accumulated by this factory over the years.

And the results?

Sample image quality is good enough for actual sampling, with far finer texture detail than the general model; when clients change requirements — color, style — images come out in minutes; it can also generate finished-product images straight from text requirements, then refine a rough draft into a polished one. Better still, different styles correspond to different weaving techniques — and so the intangible-heritage braiding methods passed down for thousands of years got "fed" into the model.

The designer went from being a "rework machine" to being "the person who reviews the images."

The Skeleton of This Playbook

Strip the whole story down and it's really three layers: use platform data to figure out market demand, use algorithms to compute the optimal product attributes, then use an AI fine-tuned on your own data to compress the design-to-production cycle to its shortest. Market-demand analysis, product development, downstream adaptation — plug and play.

This isn't a straw-hat-exclusive game. Any small manufacturer with many styles, fast iterations, and a livelihood dependent on foreign-trade orders can run the same play.

Back to that straw hat.

Craftsmanship thousands of years old, a market of 70 countries, 250 million yuan in output — and in the end it gets stuck on something as small as "wait three days for one revised image." The study was published in 2024; bring it to today, and generative AI's capabilities have climbed another notch past what they were — what took a demonstration back then would only be easier to pull off today.

What truly grinds people down has always been the small things. And what AI happens to be best at is precisely the small things.

May every straw hat dreaming of crossing the ocean be patient enough to wait for its wind.