Transforming Ad Creatives Generation for a Leading Telecommunications Company

‍

Client Background

KDDI Corporation is one of Japan’s leading telecommunications companies, headquartered in Chiyoda-ku, Tokyo. KDDI offers a wide range of mobile and fixed-line personal communication services through its au brand, as well as business solutions. As part of its efforts to implement generative AI in its operations, KDDI aims to create new value by partnering with startups that possess advanced capabilities in developing state-of-the-art AI solutions*.

*KDDI CEO Message, May 2024, https://www.kddi.com/english/corporate/kddi/president/.

‍

Challenge

KDDI promotes its services through online advertising campaigns, which must be adapted for different platforms and user behaviors. At the same time, ad creatives must adhere to strict brand guidelines, ensuring consistency in logos, colors, and messaging while staying true to the brand’s values. This process requires skilled professionals who invest significant time and effort to get it right.

‍

Although tools exist to streamline ad banner creation, they rarely leverage first-party advertising data (e.g., click-through rates). As a result, creative decisions often rely on the experience of designers and marketers rather than clear, data-driven insights, making it difficult to measure effectiveness.

‍

Recursive’s Solution

To address these challenges, Recursive, in collaboration with the KDDI group company, developed an AI-powered, data-based ad creative generation system. It enables the semi-automatic creation of ad banners while ensuring alignment with KDDI’s brand guidelines and the four visual identity styles defined for the au brand. Once the banners are generated, the system automatically selects those with the highest AI-predicted advertising effectiveness based on past campaign performance.

‍

The system’s pipeline integrates two AI models: the Image Generation Model and the KPI Prediction Model.

‍

Image Generation Model

• High-Quality Image Generation: The model leverages the latest image diffusion technology to generate realistic human images based on text descriptions while avoiding exaggerated or stereotypical elements associated with Japan that are commonly found in models trained on Western datasets (e.g., kimonos, anime-style).
• Visual Identity Customization: The model detects trigger words in prompts and applies style-specific adjustments to elements like lighting, contrast, and color. This ensures that images align with one of the au brand’s four visual identity styles—clean, friendly, playful, and advanced.
• Automatic Prompt Translation: Since the model’s API only supports English input, Recursive integrated automatic Japanese-to-English prompt translation using a large language model (LLM) to ensure seamless functionality.
• Image-to-Image Functionality: In addition to text-to-image generation, the model can modify elements of an input image—such as adding, removing, or changing accessories, clothing, hairstyles, etc.—based on the user’s prompt.
• Faster Processing: Recursive developed a request handler with asynchronous processing to efficiently manage multiple tasks. This significantly enhances response speed and ensures smooth performance for multiple users.
• Background Removal: Since the images are designed for banners, the model includes automatic background removal, which can be activated by adding a specific trigger word to the prompt. This makes post-editing easier and more efficient.

‍

‍

KPI Prediction Model

• Accurate CTR Prediction: Using a custom neural network architecture, the model predicts click-through rates (CTR) for different audience demographics (e.g., age, gender). This provides the client with data-driven insights into audience engagement and ad effectiveness.
• Real-Time Performance Estimates: The model generates predictions in less than one second, enabling the client to assess how a finalized but unreleased banner will likely perform in an ad campaign.
• High Accuracy Results: Recursive trained the model on a wide variety of historical data from different social media platforms provided by the client. This extensive training enabled the model to predict content performance levels with 72% accuracy.
• Cost-Effective Deployment: The model offers efficient processing time, even when running on standard CPU hardware, facilitating easier deployment and reducing operational costs.

‍