Evolutionary neural architecture design for text-driven computer-aided visual concept generation

Text-driven visual generation is increasingly relevant to computer-aided engineering, conceptual design and rapid visual prototyping, where alternative visual concepts must often be explored under semantic and computational constraints. However, most Generative Adversarial Network (GAN)-based text-to-image models rely on manually designed generator architectures and empirically selected text-fusion mechanisms, limiting their adaptability to different semantic generation requirements. This paper proposes TextNAS-GAN, an evolutionary neural architecture search framework for automated generator design in text-driven visual concept generation. The proposed method formulates generator discovery as a multi-objective optimisation problem that jointly considers visual fidelity, sample diversity, semantic alignment and architectural efficiency. During search, an EWSGAN-style generator supernet is optimised using NSGA-II, while a CLIP-based text–image semantic alignment objective guides architecture selection towards improved semantic alignment. A per-cell FiLM conditioning mechanism is further introduced to distribute textual modulation across multiple feature levels. To improve search reliability, the framework incorporates Top- K reevaluation, historical caching and exponential moving average-based retraining. Experiments on the fine-grained CUB-256 text-to-image generation benchmark show that the proposed method improves cross-modal semantic alignment while maintaining competitive visual quality and moderate model complexity. The results suggest that automated neural architecture design may provide a useful methodological basis for efficient and semantically controllable visual concept generation, which is relevant to future computer-aided design and engineering workflows.