--- title: "Unique Interview | Zhiyi Tech CEO: When AI Walks the Runway — What’s Next for Apparel?" author: "Unique Research" sourcePublication: "Unique Research (FFCAP English library)" originalPublishedAt: "2024-11-29T10:44:42+08:00" canonical: "https://ffcap.cn/en/research/20241129a02x6300" source: "https://ffcap.cn/en/research/20241129a02x6300" language: "en" cnSourceUrl: "https://view.inews.qq.com/a/20241129A02X6300" --- # Unique Interview | Zhiyi Tech CEO: When AI Walks the Runway — What’s Next for Apparel? _Unique Research · Unique Interview · November 29, 2024 · by Qiuping_ In 2015 Zheng Zeyu co-founded Caicloud with Google colleagues, selling AI solutions to enterprises. Working with leading MCN Ruhan, he saw apparel transforming: 2016–2017 purchase decisions shifted from traditional brands toward KOL/KOC recommendations — power structures changing, new openings. Even with Uniqlo-scale giants, apparel lacked a true monopolist. In 2018 he founded **Zhiyi Tech (知衣科技)** to pursue a platform opportunity. Statista: global fashion apparel market ~$770.9B in 2024; CAGR ~8.94% through 2029 — retail’s largest category share. Unique Research spoke with founder & CEO Zheng on how large models and GenAI empower brands and the industry. ## Self-introduction **Q:** Please introduce yourself. **Zheng Zeyu:** Computer contests from childhood; recommended to Peking University CS; CMU AI master’s; Microsoft Research, Google — always in AI. “Zhiyi” means using tech to push fashion. Since 2018 the top question is how AI helps apparel. Designers often see design as art, not tech — so AI feels puzzling. My first startup was AI outsourcing across industries. I entered apparel because design is also a _product_: demand and preference matter; data and research can target consumers and trends. Designers don’t only imagine — they research what will sell. Data and AI help them decide more rationally: market positioning and trend forecast. Working with brands, we chose to go deep: help designers find market-fit directions and create hits. That’s Zhiyi’s mission. ## What’s new with large models vs. earlier AI? **Zheng:** Tasks once impossible are feasible. **Tagging:** traditional analysis struggles to pull key info from innovative garments; large models build flexible label systems from unstructured data — even summarize trends. **Designer uplift:** once imagination alone; now AIGC auto-combines elements. Many details resist language; AI can use designer behavior, history, positioning, and persona to generate elements plus trend materials — efficiency and creativity up. ## Women’s wear: long-tail brands, copycats, high returns — can AI help? **Zheng:** What we can solve is limited — mainly (1) inventory from wrong design positioning, (2) inventory from imprecise sales allocation. Tools for deep market analysis avoid some design-mistake stock. Supply-chain linkage is harder: on our platform, spot sellable categories, produce small test lots, then quick-response scale after market proof. Needs ecosystem cooperation — we already connect 1,000+ chain firms (fabrics, garment factories) for QR capability. High women’s returns aren’t a single-tech fix (buyer vs. seller photos, sizing). Our focus: designs that sell; chain sync so we don’t over-produce early and can ramp when hot. Solving that is already a big step. ## Integration with other ops systems We have a full in-house retail ops system; supply partners who want into our ecosystem must use it. For partners who struggle, we send people to train. Apparel is fragmented — many factories have no system, resist systems, don’t see the benefit. Early stage. Shein-style: 1–2k staff running >$30B 2023 GMV is unimaginable for traditional brands; we want similar supply orchestration for high coordination. ## Challenges Integrating a long-fragmented chain: order consolidation, multi-party management, truthful data (some suppliers fake “delivered” in systems). Accuracy and timeliness are hard — system design, architecture, process, SOPs, then scale. Continuous optimization. ## GenAI acceptance Most apparel firms are interested — with designer anxiety: less grunt work is exciting; full replacement is scary. Tech gaps remain: precision and controllability of generated content; staying on style/tone; reading designer intent. We know customer needs; generation accuracy is still R&D; we watch open-source closely. ## Mid/long-tail firms lack data — problem? Even leaders rarely hold >1–3% share — SMEs and long tail dominate. We don’t train on one firm’s data; we integrate market-wide data — 100B+ apparel-related records — then personalize styles to each customer’s positioning. ## Won’t AI worsen copycatting? We see the opposite. Most designers aren’t extreme creatives — they watch peers. Without tools: scan head brands for trends, then little time left for personal exploration. Info overlap was high — same feeds, crude tags, same content for thousands of firms → homogenization. We produce millions of data points daily, distributed differentially to millions of brands. First-time teams may watch 10–20 sources; deep users grow to 1,000 / 2,000 / 5,000. Ten brands → lookalike design; 5,000 → more unique style. GenAI can push sources past 5,000 toward millions — more diversity, less sameness. ## New industry incremental space? AI strengthens differentiation — new markets, unmet demand, faster product cycles → less inventory, higher satisfaction and loyalty. Similar SKUs bore consumers; variety deepens loyalty. High-frequency updates matter more than huge stock per SKU: if every new style needs large inventory, losers sink capital. Flexibly launch 1,000 SKUs weekly/monthly with data-forecasted demand raises hit rate. Compete on refresh rate, not stock piles. ## Other GenAI reshape Broad: virtual try-on, design feedback, post-design copy and multi-style hero images, better guided selling. We prioritize production/design — style gen, smart try-on as designer copilots. Designer vs. C-end needs differ (detail vs. overall feel) — productization needs different focus per stage. ## Expectations for models in product Want smarter UX. Tools are complex; designers are intuitive — ask a question, get an answer fast. GenAI that unifies dialogue, analysis, and generation would open new distribution. Head customers out-adopt tail partly because tail lacks time/skill for complex tools — unfair. GenAI can democratize: chatbot form — ask what to study; bot asks direction, audience, status; recommends product types. Level field for every designer size. ## Chain upgrades next few years Digitize the chain with large models/GenAI. Design, production, sales data are siloed. If we integrate around design with “conservative openness” — sensitive data usable in specific scenes without broad disclosure (akin to confidential computing) — that’s a 3–5 year bet we’ll fund heavily: easier designer work + chain digitization. ## Company challenges now AIGC tech and commercial: will designers broadly accept and actually raise efficiency? Will generated designs hit or beat industry average? (Beating humans everywhere is too harsh; average is the bar.) Tech: not only image quality but measurable precision. Commercial: persuade customers to try. Balance tech and business model so AIGC ships and sticks. ## How do you define Zhiyi? Build a full industry ecosystem; tools are the means — AI, big data, cloud, whatever’s next. End goal: industry efficiency and health so every link earns better margins. First be a profitable company, then harder problems — 10, 20, 30 years to remake an industry; such chances come a few times a century. Apparel is in transition: consumers once clearly knew which brands were premium; that map is breaking; new fashion landmarks aren’t set. Fashion was top-down; consumers now want individuality — everyone can lead. Trends fragment. How do we redefine the heights, landmarks, and premiums? With the next recovery and shifting geopolitics — can we help write the next fashion history? Challenge and opportunity. _Author: Qiuping · Unique Research, November 2024_ — END — --- Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20241129a02x6300 On-site reading page: https://ffcap.cn/en/research/20241129a02x6300