AI-Driven Automated Banner Design for Visual Marketing

Issue 322

2018 Year, Issue 114

Background

In visual design, designers often spend a lot of time on simple tasks such as copy editing, basic poster layout, and multi‑size adaptation for different devices. These tasks consume 5‑6 person‑days per designer with limited impact on skill growth. At the same time, personalized marketing demands “millions of faces” for homepage banners, raising efficiency requirements. Meituan’s delivery‑tech team therefore explored AI techniques to assist designers, aiming for low‑cost, high‑efficiency, high‑quality poster generation. This article uses banner (horizontal poster) as an example to present our research on combining AI with poster design.

Analysis

What is the banner design process? We summarize it as an ordered overlay of material layers with specific visual and spatial attributes (color, shape, texture, theme, position, size, etc.). Which steps can be explored by algorithms? Prior work [1] adjusted image color distribution for magazine covers; [2] introduced saliency‑based cropping and layout optimization. Alibaba’s “Lu Ban” system generated 170 million banners on Double‑11; JD.com is developing similar systems for copy and banner design.

Figure 1: Cover color & layout design [2]

Some sub‑problems in design can be tackled algorithmically (see Figure 1). Can we build a unified learning algorithm and system to solve all banner sub‑tasks—coloring, layout, composition, generation?

Technical Solution

Material layers can be feature‑extracted and their overlay order serialized . The algorithm therefore learns “when to choose a material and where to place it”.

Figure 2: Framework

We design a planner, optimizer, and generator to form a learning‑and‑production pipeline:

Planner learns designers’ habits from data.

Optimizer refines planner output based on aesthetic quality and design principles.

Generator selects or creates materials and renders the final image.

The asset library provides material management and tagging.

Asset Library

Extracting material features is a classic classification problem. Traditional vision extracts low‑level features (color, gradient) and uses classifiers such as LR or SVM. Recent deep‑learning approaches (CNN) provide richer semantic features. We combine both low‑level and CNN‑based high‑level features for material attribute extraction.

Figure 3: Asset library – feature extraction

Planner

After data‑driven material representation, how do we learn the banner design process? Generative Adversarial Networks (GAN) are popular but have two drawbacks for banner design:

Input complexity: conditional GANs still require complex inputs (copy, target style, main content).

Output interpretability: GANs directly generate images without explicit information about material type, color, position, etc.

Because banner design is an ordered overlay of material layers, we model it with a sequence generation model . Materials are treated as “words”, the whole poster as a “sentence”. The sequence order corresponds to word order.

Figure 4: Planner – sequence generation

We train the sequence model with supervised loss at each timestep and introduce an “Object loss” (inspired by SeqGAN) to evaluate the overall sequence rationality.

Figure 5: SeqGAN

Optimizer

The planner predicts quantitative material features, but to meet aesthetic standards we apply a post‑processing optimizer. This is essentially an optimization problem . Based on design principles we define objective functions (move, scale, brightness adjustment, etc.) and use optimization methods to improve visual quality.

Figure 6: Optimizer

Generator

The optimized feature sequence is rendered by the generator. When the asset library lacks a suitable material, we employ image style‑transfer to migrate desired attributes (color, shape, texture, or high‑level visual style) from a source image to the target.

Figure 7: Material generation

Application Scenarios and Feature Extensions

“Mass personalization” is the future of marketing, demanding rich product assets and diverse poster layouts. Beyond standard style learning, we explore three extensions:

Main Image Processing

High‑quality product images require accurate segmentation. Using Fully Convolutional Networks (FCN) we adopt common techniques: Encoder‑Decoder architecture, atrous convolution, multi‑scale feature fusion, and a two‑stage fine‑tuning network.

Figure 8: Semantic segmentation & matting (based on DeepLab v3+)

Segmentation sometimes yields jagged edges. We address this by generating a trimap and applying classic matting methods (Bayesian, Closed‑Form) to obtain a smooth alpha channel.

Figure 9: Product foreground matting

We also rank source images by aesthetic quality; low‑scoring images may be enhanced via semi‑supervised GAN‑based augmentation (e.g., Cycle‑GAN).

Poster Template Expansion

Beyond learning generic styles, we use image retrieval to quickly adapt to hot scenarios. CNN and color features are extracted from assets, and Euclidean distance measures similarity, enabling automatic template expansion and poster imitation.

Figure 10: Asset retrieval & template expansion

Multi‑Resolution Expansion

Designers often need to adapt a banner to various sizes and devices. This is treated as an optimization problem: given fixed material positions, we adjust local and global layout constraints to fit any resolution within ±30% of a base aspect ratio, with plans to broaden the range.

Figure 11: Multi‑resolution adaptation

Conclusion

Our intelligent banner design system currently supports Meituan Delivery’s homepage ad slots and merchant shop decoration, with material‑processing capabilities also serving product images for flash sales. Future work includes expanding style diversity, semantic‑aware color and material mining, automatic data parsing, and building a self‑evaluating learning loop to further improve algorithmic design ability and applicability, thereby helping designers reduce repetitive work and cost.

References

[1] A. Jahanian et al., “Automatic Design of Colors for Magazine Covers”, IS&T/SPIE Electronic Imaging, 2013. [2] X.Y. Yang et al., “Automatic Generation of Visual‑Textual Presentation Layout”, ACM TOMM, 2017. [3] D.G. Lowe, “Distinctive Image Features from Scale‑Invariant Keypoints”, IJCV, 2004. [4] A. Krizhevsky et al., “ImageNet Classification with Deep Convolutional Neural Networks”, NIPS, 2012. [5] I. Goodfellow et al., “Generative Adversarial Networks”, NIPS, 2014. [6] K. Kawakami, “Supervised Sequence Labelling with Recurrent Neural Networks”, 2008. [7] T. Mikolov, “Statistical Language Models based on Neural Networks”, 2012. [8] L. Yu et al., “SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient”, AAAI, 2017. [9] L.A. Gatys et al., “Image Style Transfer Using Convolutional Neural Networks”, CVPR, 2016. [10] Y. Li et al., “A Closed‑form Solution to Photorealistic Image Stylization”, ECCV, 2018. [11] J. Long et al., “Fully Convolutional Networks for Semantic Segmentation”, CVPR, 2015. [12] L.C. Chen et al., “Encoder‑Decoder with Atrous Separable Convolution for Semantic Image Segmentation”, ECCV, 2018. [13] J.Y. Zhu et al., “Unpaired Image‑to‑Image Translation using Cycle‑Consistent Adversarial Networks”, ICCV, 2017.

Author Bio

Xiaoxing joined Meituan in June 2017, focusing on image content mining, enhancement, and generation for Meituan Delivery, aiming to translate image‑related research into production.

Join the Meituan Deep Tech Community for direct communication with project maintainers. Add WeChat ID MTDPtech02 and reply with AI to be invited automatically.