AI’s new hustle: Sell products before they’re real

Industrial deployment at Alibaba has already demonstrated the commercial viability of the system. AI-generated fashion products are live-tested on the platform, and once enough pre-orders are secured, the items proceed to manufacturing. If the minimum order threshold is not met, merchants simply cancel the production run and offer refunds.

E-commerce innovation has taken a significant leap with artificial intelligence now enabling merchants to generate and sell products before physically manufacturing them. A new system developed by researchers at Shanghai Jiao Tong University and Alibaba Group introduces AI-generated items (AIGI), allowing online sellers to create photorealistic images of personalized fashion items based on text descriptions alone. Products are only manufactured after sufficient customer demand, effectively eliminating the need for physical prototypes and reducing inventory risks.

This approach marks a fundamental shift in how e-commerce operates. Traditional product workflows involve sequential steps of design, manufacturing, photography, and marketing. These steps require high upfront costs and expose merchants to overproduction and unsold inventory. The new AI-powered process enables sellers to move directly from concept to sales, driven by consumer interest, using diffusion models that generate high-quality images without the need for real-world samples.

The researchers developed a framework called PerFusion, which uses a personalized group-level preference alignment model to generate images that not only match textual descriptions but also align with individual user preferences. The system integrates Stable Diffusion models with a custom reward mechanism, leveraging user behavior data to tailor visual content at scale. Users such as fashion designers or small store operators input text prompts describing desired product features, such as color, style, and fit, and receive a range of AI-generated image options. Based on selections, further refinements can be made until the design resonates with the intended audience.

One of the core scientific challenges addressed in the study is the variability in human visual preference. Two users may interpret the same textual prompt differently based on cultural background, shopping behavior, or aesthetic sensibility. To manage this, the system employs a reward model based on the CLIP architecture, enhanced with personalized plug-ins that learn and adapt to individual users’ tastes. These plug-ins extract behavioral patterns from user profiles and feed them into the image generation pipeline, guiding the diffusion model to produce outputs that reflect diverse consumer expectations.

The researchers evaluated the model using both public and proprietary datasets, comparing the performance of PerFusion to existing image generation techniques. Across multiple metrics, such as image quality, aesthetic alignment, and user preference satisfaction, PerFusion consistently outperformed traditional methods. In particular, it achieved a 13% relative improvement in both click-through rates and conversion rates compared to human-designed images.

Industrial deployment at Alibaba has already demonstrated the commercial viability of the system. AI-generated fashion products are live-tested on the platform, and once enough pre-orders are secured, the items proceed to manufacturing. If the minimum order threshold is not met, merchants simply cancel the production run and offer refunds. This agile manufacturing approach enables sellers to release dozens of new items per day, a feat previously constrained by the cost and time of manual design and photography.

A panel of eight industry experts, including professional fashion designers and retail specialists, conducted a blind evaluation of the AI-generated images. The AI products scored comparably to real-world items in key categories such as material representation, model realism, color coordination, and design aesthetics. The average overall score for AI-generated images reached over 94% of that of human-designed items, indicating their readiness for consumer-facing use.

The study also noted that merchants using the system experienced up to 17.8% higher click-through rates in search and 17.3% higher conversion rates. These gains are attributed to the personalization of the product visuals and the system’s ability to iterate on consumer feedback almost instantly. The model adapts based on prior shopping records, generating design variations that match the signature style of individual storefronts.

In one case study, two different merchants submitted identical prompts to the system describing a women’s slim-fit v-neck dress with long sleeves. The outputs reflected distinct brand aesthetics: one emphasized minimalism with bell sleeves, while the other favored luxury with sequins. This demonstrated that even with the same text input, the system could generate visually divergent outputs tuned to the historical preferences of each merchant.

The PerFusion system was designed to optimize not only individual image quality but also group-level performance, where users evaluate multiple generated images in context. Rather than optimizing based on individual image pairs, the model learns from user selections across entire image groups, better capturing real-world decision-making behavior.

The study positions the AI-generated item strategy as a transformative solution for sustainable and responsive e-commerce. It reduces environmental waste by eliminating overproduction, enhances product diversity through personalization, and shortens product release cycles from months to days. As the system matures, researchers anticipate broader application in product categories beyond fashion, including furniture, accessories, and digital merchandising.

The findings are published in a paper titled "Sell It Before You Make It: Revolutionizing E-Commerce with Personalized AI-Generated Items" on arXiv.