Adversarial Latent Autoencoder with Self-Attention for Structural Image Synthesis

Deep Generative Models (DGMs) have been successfully employed to synthesize general images, e.g., animals, human faces, and landscapes. This promising advancement leads to the idea of utilizing DGMs to generate novel structural designs, thereby facilitating industrial engineering processes. However, industrial design data, e.g., blueprints or engineering drawings, is fundamentally different from the images of natural scenes. They contain rich structural patterns and long-range dependencies, which are challenging for convolution-based DGMs to generate. We tackle this challenge by proposing the Self-Attention Adversarial Latent Autoencoder (SA-ALAE), which allows for generating realistic structure designs of complex engineering parts. With SA-ALAE, users can explore novel variants of an existing design and control the generation process by operating in the learned latent space. We showcase the potential of SA-ALAE by generating engineering blueprints in a real automotive design task.