Researchers' PDA-GAN AI generates ad layouts from product images with 60K dataset

A team of researchers has developed a new AI system that can automatically generate advertising poster layouts conditioned on a product image. The core challenge they address is the 'domain gap'—the difference between the inpainted poster images used for training and the clean product images used in production. To solve this, they created the Content-aware Graphic Layout Dataset (CGL-Dataset) with 60,548 annotated posters and introduced two Generative Adversarial Network (GAN) models.

Their first model, CGL-GAN, applies Gaussian blur to inpainted regions. Their second and more advanced model, the Pixel-level Discriminator Adapter GAN (PDA-GAN), incorporates unsupervised domain adaptation. PDA-GAN's key innovation is a pixel-level discriminator connected to shallow feature maps, which computes a loss for each pixel, allowing it to better adapt to the texture of clean input images and generate more coherent, image-aware layouts.

The researchers also proposed three novel content-aware evaluation metrics to assess how well a model captures relationships between graphic elements and image content. Both quantitative and qualitative evaluations demonstrated that PDA-GAN achieves state-of-the-art performance, producing higher-quality layouts that are semantically aligned with the input product image, a significant step beyond generic template-based design.