How Suning’s AI‑Powered Banner Design Platform Revolutionizes E‑Commerce Advertising

With the rise of the AI era, the clash between art and technology in commercial design has become a hot topic. Suning.com, a leader in "smart retail" that integrates online and offline shopping experiences, is actively exploring how to combine the two.

To address the inefficiency of manually creating banner ads for each product, Suning developed an intelligent design platform called "Qianbian Banner". The platform uses machine learning to generate dozens of customized banners from uploaded assets, dynamically displaying them according to user preferences and enabling personalized ad placements.

Overall Workflow and Technical Challenges

A banner consists of two main components: the product image and the decorative background (the "base"). The workflow includes:

Pre‑filtering low‑quality product images.

Using recommendation algorithms to select product images and extract the product foreground.

Generating matching base designs.

Combining product foregrounds with bases to produce multiple banners.

Scoring the banners with a model and selecting the best one.

Each stage directly impacts banner quality, and several technical difficulties arise:

Accurate segmentation of product foregrounds.

Translating design language into machine‑readable rules.

Establishing reliable scoring criteria.

Challenge 1: Product Image Segmentation

Traditional manual cut‑outs are inefficient. Suning’s solution replaces them with deep‑learning segmentation. A large pixel‑level annotated dataset trains a deep convolutional network, and dilated convolutions are used to enlarge the receptive field without pooling, preserving detail.

To refine edges, a fully connected Conditional Random Field (CRF) is applied, modeling both unary (pixel class) and pairwise (pixel similarity) terms, which sharpens object boundaries.

Anti‑aliasing is further added: the alpha channel of the segmented image is analyzed, 16 edge patterns are identified, and each pattern’s pixels are recomputed to smooth jagged edges.

Challenge 2: Converting Design Language to Machine Language

Designers provide base elements (background, texture, decorations, masks, product overlays, text). These elements are labeled by category (e.g., appliance, food) and usage (e.g., daily, promotion). A rule‑based system converts these design constraints into machine‑readable parameters, preventing incompatible combinations during base generation.

Challenge 3: Banner Scoring Model

After generating candidate banners, a multi‑task learning model evaluates them. The model jointly learns several tasks—such as product‑base compatibility, color harmony, and exposure‑click performance—using shared representations to improve generalization.

The network, inspired by Inception‑v3, replaces 3×3 convolutions with 1×3 and 3×1 kernels, adds Batch Normalization, and uses a SoftmaxWithLoss loss. Slice layers separate label inputs so each scoring item (e.g., text color, product match) receives its own prediction.

Future Directions

Three main research avenues are identified:

Adaptive layout of banner elements to support various sizes.

Intelligent coloring of base designs using designer sketches to achieve "thousand‑faces, thousand‑items" personalization.

Extending the online banner generation pipeline to offline store visual design, leveraging the same AI techniques for large‑scale retail signage.

Overall, the platform has been deployed across Suning’s online ad slots, handling massive promotional events and significantly reducing repetitive design work while maintaining high visual standards.