Smart Creative: Automated Advertising Content Generation at NetEase Yanxuan

Smartphones bring convenience but also advertising. In today's society, businesses rely heavily on advertising to attract customers, and the ads you see may be "algorithmically" customized for you.

1. Personalized Advertising

Using data analysis and algorithm technology, travel agencies can deliver different ad variations to different users. This method of accumulating advertising data, using algorithms to predict user interests, and presenting different ad content to different users is called Personalized Advertising . The advertising is not done by humans but by 7×24 hour algorithm machines.

Two possible solutions to achieve this effect: 1) Manually create multiple ad versions (costly); 2) Use basic visual materials and algorithms to automatically generate multiple ad versions (called Smart Creative ).

2. Implementation of Smart Creative

The personalized advertising system first determines the type of ad to deliver, then uses Smart Creative to synthesize ad content based on upstream information. The core is generating graphic content around products with the highest user interest.

Input: Advertising copy (main title, subtitle), Product pool (sorted by user interest)

Output: Creative content (images/videos)

2.1 Template Creation

Smart Creative cannot generate content from nothing; it needs Creative Templates to guide content generation. Templates are designed by visual experts and include abstract business elements, style definitions, and content layouts. Template elements include product main images, backgrounds, copy, decorative icons, logos, etc. (each element supports dynamic variation).

2.2 Image Synthesis

The basic approach can be described as: Creative Template + Product Image = Ad Creative. Three key problems need solving:

1. Theme Color Recognition: Scene templates use frame layouts and need to match colors based on product scene images. The first step is identifying the theme color of the product scene image to select suitable templates for rendering.

2. Aesthetic Scoring: If there are multiple candidate templates, the optimal combination needs to be selected. The approach is to first synthesize different styles of ad materials, then use algorithms to score each material, and finally select the highest-scoring material.

3. Size Matching: Multi-product composite images need to not only adapt to product sizes (avoiding poor visual experience) but also consider product diversity to make users most interested in the products included in the advertisement.

2.3 Video Synthesis

Video production costs and generation difficulty are far greater than images. The approach is to synthesize based on existing video and graphic materials: 1) Switching product hero images; 2) Adding product selling points to video materials.

2.4 Online Application

Smart Creative is not just an algorithm module but also has a complete management backend for maintaining various template information. Actual business scenarios at NetEase Yanxuan include: DSP advertising, resource position advertising, recommendations, search, etc.

3. Conclusion

Compared to offline advertising, online advertising's advantage is that the delivery data chain is trackable (exposure, click, conversion), making it easy to evaluate advertising effectiveness. With basic data and algorithm capabilities, the more refined advertising delivery, the more it can improve advertising revenue. The "creative automatic generation" technology behind this is a typical case of human experience and algorithm model collaboration, which can not only significantly reduce the cost of advertising material production but also improve advertising revenue.