How Alibaba’s AI Designer “Luban” Generates 400 Million Double‑11 Banners in Seconds

In the past, designers spent countless hours each year on Double 11, manually creating billions of posters, banners, and graphics for the massive shopping festival.

Now the AI designer “Luban” replaces that repetitive work, capable of generating up to 8,000 posters per second, or 40 million per day, aiming for 400 million banners during the event.

Luban originated from Alibaba’s desire to apply the “personalized” recommendation model to marketing graphics, leading to the establishment of the Alibaba Intelligent Design Lab. The system has evolved to a P6‑level performance within the company.

The core technology consists of three modules:

Style Learning

Luban first structures and annotates massive design assets, then trains neural networks to produce a spatial‑visual design framework. Designers label elements (e.g., product, background, masks) and define design techniques and styles, enabling the AI to understand both composition and aesthetic feel.

These annotated assets are fed into deep‑learning models that memorize complex design steps, resulting in a model of spatial and visual features that serves as a high‑level design blueprint.

Actuator

The actuator selects design prototypes from the style‑learning module and pulls appropriate elements from an element repository, planning optimal generation paths. This process mirrors a designer’s iterative adjustments and incorporates reinforcement learning to improve through trial and error.

After generating multiple design candidates, the system passes them to the evaluation network.

Evaluation Network

The evaluation network is trained on large volumes of design images and rating data, learning to judge the quality of designs. Two designer roles support Luban: one continuously updates the style‑learning module, and the other evaluates generated outputs, providing feedback for improvement.

Challenges faced during development included a lack of labeled data, the inherent uncertainty of design intent, and the absence of prior frameworks comparable to AlphaGo’s impact on Go AI.

Since its debut in 2016, Luban produced 170 million banners, doubling click‑through rates. Today it can create 40 million posters per day, each uniquely tailored to product images, and has learned from millions of design drafts.

Despite its capabilities, Luban does not replace designers; instead, it works through human‑AI collaboration, especially for innovative, creative designs that still require human insight.

Overall, Luban demonstrates controllable visual generation, turning design knowledge into data that AI can learn and apply at massive scale.