ZeroNAS presents a differentiable generative adversarial network architecture search method specifically designed for zero-shot learning (ZSL). The approach optimizes both generator and discriminator architectures, leading to significant improvements in ZSL and generalized ZSL tasks across various datasets.