AI Case Study
Alibaba Group's LuBan platform automatically designs 400M online advertising banners using image processing and reinforcement learning
The Alibaba Group's Luban platform uses machine learning for image processing to design online advertising banners for products sold on the Alibaba Group commerce sites. Its ability to generate banners at the rate of 8,000 a second allowed it to create the different ones needed to promote products for the Group's biggest online shopping day.
Consumer Goods And Services
The LuBan platform "is capable of generating approximately 8,000 different banner designs per second" using reinforcement learning and applying image processing to raw images. It is trained on human-labelled design library data and its output banners are evaluated aesthetically by humans and commercially by click-through-rates, which it then uses to learn from and improve.
Creative And Brand
In advance of Single's Day, the Alibaba Group's largest shopping day event, the LuBan AI platform was deployed to create marketing banners for millions of online stores hosted across the Group's platforms.
LuBan designed "400 million banners for an even wider range of products. If we assume it takes a human designer 20 minutes to design one single banner, then we will need 100 designers to work non-stop for 150 years to produce the same amount". This automation allowed it to create specific product advertising on a scale that would not have been possible previously.
"The machine learning algorithm LuBan employs consists of four major steps:
1. To let machine understand what is the components of design. Design team manually assigns labels to the original layers in the sourced design documents while summarizing different design styles for machine to understand.
2. To establish design element center.
After machine learns the general design frameworks, it needs a huge amount of design data to train. The team first establishes a library of design elements, then lets machine extract features from the raw images, followed by clustering these features, while inspecting the overall library quality and if there are copyright issues.
3. To let machine generate designs.
Similar to AlphaGO, the algorithm produces a virtual canvas like a Go board, then places design elements onto the board. The team employs reinforcement learning here; the machine first does some random designs, gets some meaningful feedback, then keeps iterating the whole process until it learns what kind of design is “good” or “preferred”.
4. To evaluate machine-generated designs.
Machine generated designs are evaluated from the perspective of both “aesthetics” and “commercial values”. The aesthetics valuation is done by professional agencies, while business evaluations are reflected by indicators like click-through rate."
A library of manually labelled design elements, as well as raw images